I'll produce the complete Agent Fleet Architecture document for Maik. Let me build this as a comprehensive, implementation-ready specification.

Maik Agent Fleet Architecture

Version: 1.0 Date: 16 June 2026 Author: Steve Aylward, MD, Ven Agency Status: Implementation specification

1. Architecture Overview

Maik is a premium sculptural lighting brand operated end-to-end by a fleet of fifteen named AI agents coordinated through Multica, Ven Agency's agent operating platform. The brand sells 3D-printed USB-C powered sculptural lights, manufactured on an in-house Bambu Lab print farm, fulfilled via AusPost MyPost Business, marketed on Meta and Google, and merchandised on Shopify Plus with Klaviyo as the email layer. The fleet exists to compress what would conventionally be a fifteen-person operations team into one founder plus a substrate of AI labour, while keeping brand voice, aesthetic, and customer trust at a standard a hand-built premium brand would deliver.

The architecture rests on four design principles.

Named agents, not roles. Each agent has a name, a remit, a voice, and a track record visible in Multica. This is deliberate. Anonymous agent functions blur into noise; named agents accumulate accountability, retain context across drops, and let Steve reason about the fleet as a team rather than a queue of tasks. Lumen is the photographer, not "the image generation function". Quill is the copywriter, not "Klaviyo LLM step three". Names also make handoffs legible: when Atlas hands a SKU to Lumen, the audit trail in Multica reads like a brief, not a payload.

Autonomy is earned, not granted. Every agent ships in propose-and-approve mode. Autonomy is unlocked agent by agent, surface by surface, at month 3, month 6, and month 12 based on a clean track record measured in Multica. Some agents (Smith for supplier email, Herald for press contact) never graduate to full autonomy because the downside of a misfire is too asymmetric. The ramp is encoded in the fleet's permission model, not in agent prompts, which means it survives prompt drift, model swaps, and personnel changes.

Multica is the spine, not parallel infrastructure. Every agent action, proposal, approval, escalation, and failure writes to Multica. There is no shadow log, no agent-local memory that escapes the audit trail, no Slack-only decisions. Multica holds the project graph, the approval queue, the cross-agent dependency view, and the kill switch. If Multica is down, the fleet pauses; this is intentional. The cost of running the fleet through a single substrate is that the substrate is a single point of failure. The benefit is that Steve has one place to look, not fifteen.

Fail-safe defaults. Every agent defaults to the action that minimises customer-facing damage. Atlas defaults to draft, not publish. Beacon defaults to budget pause, not spend increase. Echo defaults to escalate to Steve, not auto-reply. Anvil defaults to pause the print queue, not push a job. The cost of over-cautious defaults is a longer founder review queue. The benefit is that no agent can silently incinerate the brand at 3am.

Forge as orchestrator. Forge is the conductor. It does no operational work itself — it does not generate images, write copy, place ad bids, or contact suppliers. Forge routes work between agents, sequences multi-agent workflows, arbitrates priority conflicts, composes the Monday digest, holds the approval gate, and surfaces escalations to Steve. Forge's value is that the other fourteen agents can stay narrow and excellent at their craft while Forge handles the connective tissue.

Shared versus per-brand agents. Seven agents (Atlas, Anvil, Hearth, Smith, Oracle, Echo, Mercury) are operationally identical across any future Ven-operated brand. Their value scales super-linearly: building them once and pointing them at a new Shopify store costs near-zero marginal effort. Five agents (Quill, Lumen, Pulse, Herald, Ember) are brand-differentiation agents — they encode Maik's voice, visual identity, and editorial sensibility. These scale sub-linearly: each new brand requires a fresh voice-adapter fine-tune, a fresh visual library, a fresh editorial calendar. Two agents (Beacon, Sage) sit between — the mechanics are shared, the creative briefs are brand-specific.

Steve's interface. Steve interacts with the fleet through three surfaces and three surfaces only. (1) A Monday 7am AEST digest from Forge summarising the prior week, anomalies, and the queue ahead. (2) A live approval queue in Multica, ordered by urgency, with one-click approve/reject/edit. (3) A single Slack thread per active project, mirrored in Multica, for ambient context and ad-hoc questions. Every other agent surface is a tool, not an interface — Steve does not log into Shopify Admin, Meta Ads Manager, or Klaviyo to operate the brand. He logs in to verify when Forge flags something worth verifying.

2. Orchestrator — Forge

Role. Forge is the conductor of the Maik fleet. It owns no operational surface, holds no API keys to customer-facing systems beyond Multica itself, and produces no asset, copy, ad, or email. Forge exists to sequence the other fourteen agents, hold the approval gate, compose the Monday digest, arbitrate priority conflicts, and route escalations. Treating Forge as a peer that "does orchestration work" is a category error; Forge is the substrate the other agents run on top of.

Responsibilities.

Maintain the project graph in Multica — every active workflow (SKU launch, drop, campaign, incident) is a project node with dependencies between agents.
Sequence multi-agent workflows according to the four canonical patterns in Section 4.
Hold the approval gate. When any agent surfaces a propose-and-approve action, Forge writes it to Steve's queue, attaches context, and blocks downstream work until Steve responds.
Compose the Monday 7am AEST digest from Oracle's analytics roll-up plus the project graph state.
Arbitrate priority conflicts — when Hearth wants to pause a SKU for inventory reasons and Beacon wants to scale spend on the same SKU, Forge holds, surfaces the conflict to Steve, and proposes a default resolution.
Route escalations from any agent to Steve, with a structured incident card.
Run the daily 06:00 AEST health check across all fourteen agents and surface any agent that has failed, drifted, or been silent beyond its expected cadence.
Maintain the autonomy ledger — which agent has which permission at which surface, gated by the month-3/6/12 ramps.

Autonomy boundary. Forge is always propose-and-approve at the boundary with Steve. Internal to the fleet, Forge is autonomous: it can sequence, route, hold, and arbitrate without per-action approval. Forge's autonomy never expands at month 3, 6, or 12 because Forge does not operate on customer-facing surfaces. Its only privileged action is composing the digest and managing the queue, both of which Steve reads before acting.

Integrations. Forge speaks only to Multica. Every other agent reports to Forge through Multica project rows, issue threads, and event subscriptions. Forge does not hold credentials for Shopify, Meta, Google, Klaviyo, AusPost, Higgsfield, or Bambu/OctoEverywhere. This is deliberate — Forge cannot accidentally take a customer-facing action.

Cadence. Always-on. Forge processes Multica events in real time, runs the 06:00 health check daily, composes the digest weekly (Monday 06:30 AEST, delivered 07:00), and runs a fleet retrospective monthly on the last Friday.

Escalation rules. Forge escalates to Steve immediately on: agent failure beyond retry budget; cross-agent priority conflict that cannot be resolved by policy; any agent attempting an action outside its autonomy boundary; Multica health degradation; external integration outage longer than 30 minutes during business hours. Escalations write to the top of Steve's queue with a red flag and a Slack ping.

Failure modes. (1) Forge could compose a digest that misses a critical anomaly because Oracle's roll-up was incomplete — mitigated by Forge cross-checking the project graph for stale-but-active issues. (2) Forge could route a workflow incorrectly, sending a SKU to Quill before Lumen finished photography — mitigated by hard dependency edges in the project graph that Forge cannot bypass. (3) Forge could deadlock on a priority conflict — mitigated by a default-resolution policy where, after 4 hours unresolved, Forge takes the conservative path (pause, do not act) and escalates.

Recovery patterns. On any agent failure, Forge first retries with exponential backoff (three attempts over 15 minutes), then pauses the project and writes an incident card. On Multica degradation, Forge enters read-only mode — the fleet stops taking new actions but continues to log state. On its own failure, Forge is restarted from the last consistent Multica snapshot; the project graph is the source of truth.

3. Agent Roster — 14 Specialists

3.1 Atlas — Listings and Catalogue

Role and remit. Atlas owns the Shopify product catalogue. It creates new product records, maintains variant structures, handles inventory sync with Hearth, manages metafields, ensures every SKU has the correct merchandising attributes (collection, tag, vendor, type), and acts as the source of truth for "what is currently sellable on Maik.com". Atlas is the first agent in every new-SKU workflow because no other agent can do useful work until a product record exists.

Specific responsibilities.

Create draft Shopify product records when a new SKU graduates from the design pipeline.
Maintain variant matrices (size, finish, USB-C cable colour, plate variant).
Sync inventory levels with Hearth via the Shopify Inventory API.
Manage product metafields for structured data (material composition, dimensions, print time, weight, recommended room).
Maintain collection membership and merchandising order on collection pages.
Apply and rotate seasonal tags for filtering and search.
Run a daily catalogue audit: every product has at least four images, copy, metafields, inventory state, and a price. Flag any gaps to Forge.
Manage product retirement — when a SKU is sunset, Atlas drafts the redirect, archives the product, and notifies Sage and Beacon.
Maintain the canonical SKU schema in Multica that all other agents reference.
Handle bulk price updates with a propose-and-approve gate.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: new product creation, price changes, collection structure changes, product retirement. AUTONOMOUS on: inventory level sync, metafield maintenance, daily catalogue audit, tag rotation within an approved tag list. HUMAN-ONLY on: launching a new collection structure or category.
Month 3-6. Adds AUTONOMOUS on: product creation in draft state (still gated for publish), routine metafield updates, image swap-ins from Lumen-approved library.
Month 6-12. Adds AUTONOMOUS on: publish for product updates (not new SKUs), price changes within 10% of last approved price.
Month 12+. AUTONOMOUS on: new SKU publish for products matching an existing template with full creative pipeline complete. PROPOSE-AND-APPROVE remains on: new collections, category expansion, sitewide price changes.
Indefinite human-only. Sitewide repricing, brand category expansion (e.g., adding non-lighting), removal of an entire collection.

MCP integrations. Shopify Admin API (products, variants, inventory, collections, metafields, redirects), Multica (project rows, audit log), Hearth (inventory feed), Lumen (image library), Quill (copy slots), Sage (URL and redirect coordination).

Cadence. Always-on for inventory sync and webhooks. Daily catalogue audit at 05:00 AEST. Event-triggered on new SKU graduation from design.

Approval queue feed. New product proposals surface to Steve as a single card with proposed handle, title, price, variant matrix, collection assignment, and a "publish on" date. SLA: 24 hours during drop windows, 72 hours otherwise. Price changes surface as a batch weekly with old/new/delta and rationale from Beacon or Oracle.

Escalation rules. Escalates immediately on Shopify API failure that prevents inventory sync, any product with a stockout that Hearth has not pre-flagged, or a 5xx rate above 1% over a 10-minute window.

Failure modes and safeguards. (1) Atlas could publish a SKU before Lumen has uploaded final imagery — safeguarded by Forge's dependency graph blocking publish until all four pipeline agents have signed off. (2) Atlas could drift inventory out of sync with the print farm — safeguarded by twice-daily reconciliation with Hearth's queue depth and Bambu's print log. (3) Atlas could orphan a redirect — safeguarded by Sage running a post-action redirect audit within 1 hour of any product retirement.

3.2 Lumen — Photography and Imagery

Role and remit. Lumen produces all product imagery and lifestyle imagery for Maik. It generates hero shots, alternate angles, scale-reference shots, lifestyle compositions, and seasonal campaign imagery using Higgsfield as its primary generation engine, with a curated reference library and a Maik-specific style adapter. Lumen is the visual voice of the brand and the most tightly gated of the differentiation agents, because off-brand imagery is the fastest way to erode a premium positioning.

Specific responsibilities.

Generate product imagery for every new SKU: minimum four shots (hero, three-quarter, detail, scale).
Generate lifestyle imagery for collection pages and EDM hero modules.
Maintain the Maik visual reference library in Multica — every approved image is tagged with palette, light condition, room context, and use case.
Run weekly visual drift checks — sample ten recent generations and score against the Maik style guide. Flag drift above 15%.
Brief Higgsfield prompts using a structured template that encodes the Maik palette (paper, ink, eucalyptus, brass), light language (warm, low, sculptural), and composition rules.
Handle reference-image-to-final-image variation runs for A/B testing on Meta and EDM.
Coordinate with Atlas on which images are attached to which product variant.
Coordinate with Beacon on ad-creative variants and Pulse on social-native crops.
Maintain the seasonal campaign moodboard ahead of each drop.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: every image attached to a product, every hero image, every EDM hero. AUTONOMOUS on: internal moodboard drafting, reference library tagging, drift checks. HUMAN-ONLY on: brand visual identity changes.
Month 3-6. Adds AUTONOMOUS on: variant generation from an approved hero (e.g., colour swap, crop variations). Voice-adapter fine-tune ships by end of month 3.
Month 6-12. Adds AUTONOMOUS on: lifestyle imagery generation against a pre-approved brief, ad-creative variants under Beacon's brief.
Month 12+. AUTONOMOUS on: full SKU imagery pipeline for products matching an existing template. PROPOSE-AND-APPROVE remains on: hero campaign imagery, new visual territories, seasonal moodboard.
Indefinite human-only. Brand visual identity changes, founder portrait or human-figure imagery, any image used in press.

MCP integrations. Higgsfield (generation), Multica (asset library, approval queue), Shopify Admin API (image upload via Atlas), Meta Ads API (creative library via Beacon), Klaviyo (asset library via Ember).

Cadence. Event-triggered on new SKU. Drop-triggered for campaign work. Weekly drift check on Friday.

Approval queue feed. Hero imagery surfaces as a grid of four candidates with the moodboard reference and Lumen's recommended pick. SLA: 48 hours during drop windows, 5 days otherwise. Variant runs surface as batches.

Escalation rules. Escalates on drift score above 15% across a single batch, Higgsfield API outage, or any image that triggers the brand-safety classifier (figure rendering, brand logos other than Maik's, anything resembling competitor products).

Failure modes and safeguards. (1) Lumen could publish off-brand imagery — safeguarded by mandatory Steve approval through month 6 on every product image, and by drift checks thereafter. (2) Lumen could exhaust the Higgsfield budget on variant runs — safeguarded by a monthly generation budget tracked in Multica with Forge pausing at 80% utilisation. (3) Lumen could generate imagery that infringes on a reference style — safeguarded by the prompt template excluding style references to living artists or named studios.

Voice and brand discipline. Lumen ships with a Maik visual style guide encoded as a structured prompt prefix plus a curated reference library of 200 approved images by month 3. The voice-adapter fine-tune ships by end of month 3 and is retrained quarterly against the growing approved library. Drift checks compare new generations to the most recent 50 approved images on palette, composition, and light language.

3.3 Quill — Copywriting

Role and remit. Quill writes everything Maik says in text. Product descriptions, collection page copy, EDM body copy, ad copy, social captions, blog posts, FAQ entries, and post-purchase emails. Quill is the verbal voice of the brand — restrained, tactile, never-shouty, never-jargonny, never-cute. It is the second-most-gated differentiation agent after Lumen because off-brand voice is the second-fastest way to erode premium positioning.

Specific responsibilities.

Write product descriptions to a structured template: hero line, three-paragraph descriptive body, materials block, dimensions block, care block.
Write collection page narratives — one to three short paragraphs that frame the collection.
Write EDM copy in coordination with Ember, including subject lines, preview text, hero copy, and section bodies.
Write Meta and Google ad copy variants in coordination with Beacon.
Write social captions in coordination with Pulse.
Maintain the Maik voice guide in Multica — tone rules, banned phrases, preferred constructions, sentence-length distribution.
Run weekly voice drift checks across published copy.
Write post-purchase email sequences in coordination with Ember.
Write FAQ entries triggered by Echo flagging repeat customer questions.
Handle copy localisation when the brand expands internationally (post month 12).

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: every customer-facing string. AUTONOMOUS on: internal drafting, voice guide maintenance, drift checks. HUMAN-ONLY on: brand voice guide changes.
Month 3-6. Adds AUTONOMOUS on: ad copy variants under Beacon's brief, FAQ entry drafting, social caption variants. Voice-adapter fine-tune ships by end of month 3.
Month 6-12. Adds AUTONOMOUS on: product description publishing for SKUs matching an existing template, EDM body copy under Ember's brief.
Month 12+. AUTONOMOUS on: most customer-facing copy with weekly batch review. PROPOSE-AND-APPROVE remains on: brand-defining moments (drop announcement, collection launch, About page), press statements.
Indefinite human-only. Brand voice guide changes, About page, founder voice, press statements, crisis communications.

MCP integrations. Shopify Admin API (product descriptions via Atlas), Klaviyo (campaign body via Ember), Meta Ads API (ad copy via Beacon), Multica (voice guide, drift checks).

Cadence. Event-triggered on new SKU, drop, EDM, ad campaign. Weekly drift check on Friday.

Approval queue feed. Copy surfaces as a structured card with the brief, three variants, and Quill's recommended pick. SLA: 24 hours during drop windows, 72 hours otherwise.

Escalation rules. Escalates on voice drift above 20% in any batch, banned-phrase detection, or any copy referencing claims (warranty, certification, origin) that Steve has not verified.

Failure modes and safeguards. (1) Quill could drift toward generic e-commerce voice — safeguarded by drift checks and a banned-phrase list ("elevate your space", "must-have", "game-changer"). (2) Quill could fabricate product claims — safeguarded by a claims allowlist in Multica; any claim outside the list is auto-flagged. (3) Quill could produce inconsistent copy across surfaces — safeguarded by Forge cross-checking product copy against ad copy against EDM copy at drop time.

Voice and brand discipline. Quill ships with a structured voice guide (sentence length 8 to 22 words, two-syllable bias, no exclamation marks, no second-person imperative outside CTAs, no superlatives without specification). A voice-adapter fine-tune on 500 approved copy samples ships by end of month 3 and is retrained quarterly.

3.4 Beacon — Paid Media

Role and remit. Beacon operates Meta Ads and Google Ads for Maik. It builds campaigns, manages bids and budgets, rotates creative, analyses performance, proposes optimisations, and reports to Oracle for cross-channel attribution. Beacon is operationally aggressive but financially conservative — defaults always favour pausing spend rather than scaling it.

Specific responsibilities.

Build Meta campaign structures aligned with Andromeda best practice — consolidated ad sets, broad audiences, creative diversity within ad sets.
Build Google campaigns covering Performance Max for shopping, Search for branded and high-intent terms, and YouTube for video assets when Lumen produces them.
Manage daily budgets with a hard ceiling set by Steve.
Rotate creative on a fatigue-detection trigger (frequency above 2.5, CTR decay above 30%).
Run weekly creative-performance review and brief Lumen and Quill on next-cycle variants.
Pause underperformers — ads, ad sets, campaigns — against pre-agreed thresholds.
Maintain Meta Pixel and Google Ads conversion tagging in coordination with GTM and Oracle.
Run weekly anomaly detection on spend, CPA, ROAS, CTR.
Coordinate drop launches with Herald and Ember on launch-day ramp schedules.
Maintain a creative testing roadmap with hypotheses, sample sizes, and verdicts.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: campaign creation, budget changes above 10%, new audience definition, creative go-live. AUTONOMOUS on: pause actions for spend safety, bid adjustments within 15%, fatigue rotations within an approved creative pool. HUMAN-ONLY on: monthly budget envelope, new platform onboarding.
Month 3-6. Adds AUTONOMOUS on: budget changes within 20%, new ad sets within an existing campaign, audience refinement within an existing strategy.
Month 6-12. Adds AUTONOMOUS on: campaign creation for established product categories, budget changes within 30%, creative go-live for variants of an approved master creative.
Month 12+. AUTONOMOUS on: most campaign operations within the monthly envelope. PROPOSE-AND-APPROVE remains on: monthly budget envelope, new platform expansion, brand campaigns.
Indefinite human-only. Monthly budget envelope, platform expansion (TikTok, Pinterest), brand campaigns vs performance campaigns split.

MCP integrations. Meta Ads API (Marketing API v21.0), Google Ads API (v17+), GA4 (via Oracle), GTM (conversion tag health), Multica (campaign log, anomaly feed).

Cadence. Always-on for monitoring. Daily 09:00 AEST performance check. Weekly Monday strategy review. Event-triggered on drop launch and anomaly.

Approval queue feed. Campaign launches surface as a structured card with audience, budget, duration, creative attached, and expected CPA range. SLA: 48 hours pre-drop, 5 days otherwise. Anomalies surface immediately with proposed action.

Escalation rules. Escalates immediately on: spend pace 25% above expected, CPA 50% above baseline for 24 hours, ROAS below 1.5 for 48 hours on any campaign with material spend, ad disapproval at the account level, account-level frequency above 3.5.

Failure modes and safeguards. (1) Beacon could overspend on a poorly performing campaign — safeguarded by hard daily budget caps in Meta/Google plus a Beacon-side pause trigger at 80% of monthly envelope. (2) Beacon could approve creative that fails platform policy — safeguarded by Quill and Lumen running pre-flight policy checks. (3) Beacon could chase short-term CPA at the expense of brand equity — safeguarded by Oracle's monthly brand-search lift report and Steve's monthly creative review.

3.5 Sage — SEO

Role and remit. Sage owns organic search performance for Maik. It maintains technical SEO health, runs content gap analysis, manages internal linking, monitors keyword rankings, handles redirects in coordination with Atlas, and briefs Quill on content opportunities. Sage is patient — it operates on monthly horizons, not daily.

Specific responsibilities.

Run weekly technical SEO audits — indexation, sitemap health, Core Web Vitals, structured data.
Maintain the keyword universe in Multica — head terms, long-tail, branded, competitor.
Monitor rankings via Google Search Console and DataForSEO.
Brief Quill on monthly content priorities based on keyword opportunity and SERP feature analysis.
Manage internal linking across collection pages, product pages, and blog posts.
Coordinate with Atlas on URL structure and redirects.
Maintain product structured data (Product schema, Offer schema, AggregateRating schema as reviews mature).
Manage robots.txt, sitemap.xml, hreflang (when international launches).
Monitor backlink profile via DataForSEO and flag toxic links.
Run monthly competitor SEO analysis.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: new content pieces, URL structure changes, robots.txt changes, sitemap structure changes, redirect maps. AUTONOMOUS on: monitoring, audits, internal linking within an existing structure, structured data on new products. HUMAN-ONLY on: site architecture changes.
Month 3-6. Adds AUTONOMOUS on: minor metadata updates, schema additions, sitemap submissions.
Month 6-12. Adds AUTONOMOUS on: blog post publishing where Quill has produced approved copy, URL structure changes for product retirement.
Month 12+. AUTONOMOUS on most SEO operations. PROPOSE-AND-APPROVE remains on: site architecture changes, new market launch SEO setup, canonical strategy changes.
Indefinite human-only. Site architecture changes, canonical strategy, new market SEO foundation.

MCP integrations. Google Search Console API, DataForSEO API, Shopify Admin API (via Atlas), GA4 (via Oracle), Multica.

Cadence. Weekly technical audit on Monday. Monthly content brief on first of month. Daily ranking monitor.

Approval queue feed. Content briefs surface monthly as a batch. Technical changes surface as individual cards. SLA: 5 days for content briefs, 48 hours for technical changes.

Escalation rules. Escalates on indexation drop above 10%, Core Web Vitals regression, ranking drop above 5 positions on head terms, structured data errors blocking rich results.

Failure modes and safeguards. (1) Sage could approve content that cannibalises existing rankings — safeguarded by pre-publish cannibalisation check against the keyword universe. (2) Sage could break the sitemap — safeguarded by post-deploy sitemap validation and Search Console error monitoring. (3) Sage could chase low-quality keywords — safeguarded by an opportunity score that weights search volume, intent fit, and SERP difficulty.

3.6 Ember — EDM and Klaviyo

Role and remit. Ember runs Klaviyo for Maik. It manages campaigns, flows, segmentation, deliverability, and revenue attribution. Ember coordinates with Quill on copy and Lumen on imagery. It is brand-disciplined because email is the highest-trust channel and the easiest to over-fish.

Specific responsibilities.

Build and send drop-launch campaigns, post-drop follow-ups, and weekly nurture sends.
Maintain core flows: welcome (4-step), abandoned cart (3-step), browse abandonment (2-step), post-purchase (3-step), winback (3-step).
Manage segmentation: engaged, lapsed, high-value, first-time, repeat.
Monitor deliverability — bounce rate, complaint rate, sender reputation.
Run A/B tests on subject lines, send times, hero modules.
Maintain the email design system in Klaviyo template form.
Run weekly list hygiene — suppress hard bounces, flag suspicious sign-ups.
Coordinate with Beacon on retargeting suppression of recent purchasers.
Maintain SMS marketing as a future module (post month 6).
Report attributed revenue weekly to Oracle.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: every send, flow change, segment change, template change. AUTONOMOUS on: scheduled sends previously approved, list hygiene, deliverability monitoring. HUMAN-ONLY on: brand sender setup, domain authentication, list import.
Month 3-6. Adds AUTONOMOUS on: send execution on approved sends (Steve approves the campaign, Ember executes scheduled sends), flow optimisation within approved flows.
Month 6-12. Adds AUTONOMOUS on: campaign creation within approved templates, segment refinement, A/B test execution.
Month 12+. AUTONOMOUS on: standard nurture and product campaigns. PROPOSE-AND-APPROVE remains on: drop-launch campaigns, brand moments, list import or export.
Indefinite human-only. List import or export, domain or sender authentication changes, segments crossing personal data sensitivity thresholds.

MCP integrations. Klaviyo API, Shopify Admin API (purchase data sync), Multica.

Cadence. Always-on for flows. Weekly campaign send (Thursday 11am AEST default). Drop-triggered campaigns. Daily deliverability check.

Approval queue feed. Campaigns surface as a structured card with audience, send time, subject line variants, preview, attributed-revenue projection. SLA: 48 hours pre-send.

Escalation rules. Escalates on complaint rate above 0.1%, bounce rate above 2%, deliverability score drop, or any send to a segment above 50,000 contacts.

Failure modes and safeguards. (1) Ember could blast an unsegmented list — safeguarded by mandatory segment size confirmation above 10,000 contacts. (2) Ember could send during deliverability degradation — safeguarded by a pre-send health check that pauses sends if reputation drops below threshold. (3) Ember could send copy off-voice — safeguarded by Quill's voice check on every send.

Voice and brand discipline. Ember inherits Quill's voice guide and Lumen's visual library. Send cadence is capped at one campaign per week steady state, scaling to three in drop weeks. Drop fatigue tracked by Oracle.

3.7 Anvil — Fulfilment Scheduling

Role and remit. Anvil orchestrates the print farm and fulfilment pipeline. When an order comes in, Anvil decides which printer prints which part, in which order, with which filament, and when. Anvil coordinates with Hearth on inventory and Smith on supplier reorders for filament.

Specific responsibilities.

Subscribe to Shopify orders/create webhook and write a print job row to the Multica print-queue table with priority calculated as: (ship-by date − current time) × order value × customer LTV multiplier.
Allocate print jobs to specific Bambu printers via OctoEverywhere, balancing queue depth, filament colour change cost, and printer health.
Schedule print start times to minimise filament changes (batch like-colour jobs).
Monitor print progress via OctoEverywhere — flag stalls, failures, and bed-clearing needs.
Generate pick-pack labels via AusPost MyPost Business once a print completes successfully.
Update Shopify order status (in production, ready to ship, shipped) and send tracking.
Coordinate with Hearth on filament consumption forecasting.
Run nightly schedule optimisation for the next 72 hours of queue.
Handle reprint logic for failed prints with automatic priority bump.
Maintain a print quality log linked to each completed order.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: schedule changes affecting more than 10 orders, escalations to expedited shipping. AUTONOMOUS on: print job allocation, schedule optimisation, label generation, order status updates, reprint of failed prints. HUMAN-ONLY on: printer fleet changes, filament brand changes.
Month 3-6. Adds AUTONOMOUS on: expedited shipping for orders meeting pre-agreed criteria.
Month 6-12. Adds AUTONOMOUS on: queue prioritisation overrides, schedule changes within the 72-hour horizon.
Month 12+. AUTONOMOUS on full fulfilment operations. PROPOSE-AND-APPROVE remains on: printer fleet additions, filament brand changes, fulfilment SLA changes.

MCP integrations. Shopify Admin API (orders, fulfilments), OctoEverywhere API (print queue, printer status), Bambu Studio (via OctoEverywhere), AusPost MyPost Business API (labels, tracking), Multica (print queue table, fulfilment log).

Cadence. Always-on event-driven. Nightly 02:00 AEST schedule optimisation. Daily 08:00 AEST fulfilment digest.

Approval queue feed. Surfaces escalations only — expedited shipping requests, major schedule disruptions. SLA: 4 hours during business hours.

Escalation rules. Escalates on print failure rate above 5% over 24 hours, printer offline above 2 hours, AusPost API failure, ship-by SLA breach risk above 10% of orders.

Failure modes and safeguards. (1) Anvil could over-allocate a printer — safeguarded by per-printer queue depth caps and health checks. (2) Anvil could ship the wrong order — safeguarded by post-print QR-code verification at pick-pack. (3) Anvil could miss a ship-by date silently — safeguarded by SLA breach alerts at 24h, 12h, and 6h before deadline.

3.8 Hearth — Inventory and Supplier Reorder

Role and remit. Hearth maintains inventory state across raw materials (filament, USB-C cables, packaging, plates) and finished goods (the rare case where Maik holds finished stock). Hearth forecasts demand, calculates reorder points, drafts purchase orders, and hands them to Smith for supplier contact.

Specific responsibilities.

Track filament inventory by colour, brand, and lot.
Track USB-C cable inventory by variant.
Track packaging inventory (boxes, void fill, branded inserts).
Forecast 30-day demand by SKU using Oracle's sales velocity data.
Calculate reorder points incorporating lead time and safety stock.
Draft purchase orders when reorder points hit.
Hand POs to Smith for supplier outreach.
Reconcile received inventory against POs.
Maintain BOM (bill of materials) for every SKU.
Run weekly inventory variance reports comparing system to physical counts.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: every purchase order, BOM changes, reorder point changes. AUTONOMOUS on: forecasting, variance reporting, inventory tracking. HUMAN-ONLY on: supplier additions, supplier removals.
Month 3-6. Adds AUTONOMOUS on: POs to existing suppliers below $1,000.
Month 6-12. Adds AUTONOMOUS on: POs to existing suppliers below $3,000, reorder point adjustments within 20%.
Month 12+. AUTONOMOUS on POs below $5,000 to established suppliers. PROPOSE-AND-APPROVE remains on: new suppliers, large POs, BOM changes.

MCP integrations. Shopify Inventory API (via Atlas), supplier portals where APIs exist (otherwise via Smith), Multica.

Cadence. Daily 06:00 AEST inventory check. Weekly forecast refresh. Event-triggered on stock thresholds.

Approval queue feed. POs surface as structured cards with supplier, items, quantities, cost, expected delivery. SLA: 48 hours.

Escalation rules. Escalates on projected stockout within 7 days, supplier delivery slip above 5 days, variance above 5% on weekly count.

Failure modes and safeguards. (1) Hearth could under-forecast a viral SKU — safeguarded by Oracle's anomaly detection feeding Hearth on velocity spikes. (2) Hearth could double-order — safeguarded by Multica PO de-duplication checks. (3) Hearth could lose inventory drift — safeguarded by weekly physical count reconciliation.

3.9 Smith — Supplier Communication

Role and remit. Smith handles all written communication with suppliers — filament vendors, packaging suppliers, hardware suppliers, freight forwarders. Smith drafts emails, negotiates terms within agreed bands, schedules calls, and maintains the supplier relationship history. Per the organisation-level rule, Smith never sends outbound email without Steve's explicit approval. This constraint is permanent.

Specific responsibilities.

Draft purchase order emails from Hearth's POs.
Draft follow-up emails on late deliveries.
Draft negotiation emails on pricing, MOQ, lead time.
Maintain supplier contact records and relationship history.
Track supplier performance — on-time delivery, quality, response time.
Draft RFQs for new categories.
Coordinate with Anvil on inbound delivery scheduling.
Maintain supplier contracts and renewal schedule in Multica.

Autonomy boundary.

Month 0-3 through Month 12+. PROPOSE-AND-APPROVE on all outbound email indefinitely. AUTONOMOUS on: drafting, supplier record maintenance, performance tracking, contract calendar. HUMAN-ONLY on: outbound send.
Indefinite human-only. Outbound email send. This is locked by org rule and does not graduate.

MCP integrations. Gmail (draft only, no send), Multica (supplier records, draft queue), Hearth (PO feed).

Cadence. Event-triggered by Hearth POs and supplier replies. Weekly supplier performance digest.

Approval queue feed. Every draft email surfaces in the approval queue. Steve reviews and presses send from his own inbox or approves Smith to send via Hermes (a future identity carrier; through month 12 Steve sends from his own inbox). SLA: 24 hours.

Escalation rules. Escalates on supplier non-response beyond 5 business days, supplier price changes above 10%, quality issues from received inventory.

Failure modes and safeguards. (1) Smith could draft an email that commits Maik to terms Steve has not approved — safeguarded by the indefinite approval gate. (2) Smith could lose context on supplier history — safeguarded by the Multica supplier-record system being the only allowed source of relationship truth. (3) Smith could draft against an outdated PO — safeguarded by drafts auto-expiring after 24 hours if not approved.

3.10 Echo — Customer Service Triage

Role and remit. Echo is the first responder to every customer message. It reads, classifies, drafts responses, surfaces patterns, and routes complex cases to Mercury or Steve. Echo never sends customer email without approval through month 6 (and brand-critical replies remain gated indefinitely).

Specific responsibilities.

Read every inbound email, contact form, social DM, and live chat message.
Classify each message by type: order status, product question, return request, complaint, press, vendor pitch, spam.
Draft a response from a library of approved templates.
Surface drafts to Steve for approval (month 0-3) or send approved templates directly (month 6+).
Route returns to Mercury, press to Herald, vendor pitches to Steve.
Maintain the FAQ library and feed Quill on repeat questions worth publishing.
Run weekly sentiment analysis on customer messages and report to Oracle.
Maintain a customer-context view — order history, prior interactions, segment.
Flag VIP customers (high LTV, press contacts) for white-glove handling.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: every customer reply. AUTONOMOUS on: classification, routing, FAQ updates, sentiment analysis. HUMAN-ONLY on: complaint escalations, press contacts.
Month 3-6. Adds AUTONOMOUS on: replies using exact-match approved templates for order-status and standard product questions.
Month 6-12. Adds AUTONOMOUS on: replies using approved templates with light personalisation, return-initiation responses (routed to Mercury).
Month 12+. AUTONOMOUS on: most routine replies. PROPOSE-AND-APPROVE remains on: complaints, refunds outside policy, brand-sensitive messages.
Indefinite human-only. Press inquiries, legal/regulatory inquiries, crisis comms.

MCP integrations. Gmail (read and draft, send only for approved templates after month 6), Shopify Admin API (customer context), Klaviyo (customer profile via Ember), Multica.

Cadence. Always-on. Hourly digest of pending replies. Daily sentiment report.

Approval queue feed. Replies surface in batches three times daily (09:00, 13:00, 17:00 AEST). Urgent items (complaints, press) surface immediately. SLA: 4 hours during business hours.

Escalation rules. Escalates immediately on press contact, legal threat, complaint with public-channel risk (e.g., customer mentioning social media), refund above $200, VIP customer message.

Failure modes and safeguards. (1) Echo could misclassify a press inquiry as routine — safeguarded by a press-detection classifier with high recall and Steve approval on anything ambiguous. (2) Echo could send an off-policy template — safeguarded by the template library being the only allowed source of approved language. (3) Echo could miss a complaint — safeguarded by daily zero-inbox audit by Forge.

3.11 Mercury — Returns and Refunds

Role and remit. Mercury handles the returns and refunds workflow. When Echo routes a return request, Mercury validates eligibility, issues return labels, tracks the inbound shipment, processes the refund, and updates inventory via Hearth.

Specific responsibilities.

Validate return eligibility against policy (30-day window, condition requirements).
Generate AusPost return labels via MyPost Business.
Track inbound return shipments.
Inspect received returns (via photograph workflow from receiving) and classify as resalable, refurbishable, or scrap.
Process refunds via Shopify.
Update Hearth on returned inventory.
Maintain return reason taxonomy and feed Oracle on quality signals.
Handle warranty claims for products within warranty window.
Draft refund-confirmation messages for Echo to send.
Flag patterns (e.g., a SKU with elevated return rate) to Forge.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: every refund, every return-label issuance above standard policy. AUTONOMOUS on: eligibility validation, label generation within policy, inbound tracking, inventory updates. HUMAN-ONLY on: refunds above $300, warranty claims.
Month 3-6. Adds AUTONOMOUS on: refunds within policy below $150.
Month 6-12. Adds AUTONOMOUS on: refunds within policy below $300, exchange processing.
Month 12+. AUTONOMOUS on: full returns within policy. PROPOSE-AND-APPROVE remains on: refunds outside policy, warranty claims, partial refunds as compensation.

MCP integrations. Shopify Admin API (refunds, orders), AusPost MyPost Business API (return labels), Multica.

Cadence. Event-triggered on return request. Daily 17:00 AEST returns digest.

Approval queue feed. Refunds outside policy surface as cards with customer history, return reason, requested amount, recommended action. SLA: 24 hours.

Escalation rules. Escalates on return rate by SKU above 5%, fraud signals (multiple returns from same customer or address), warranty claims, refund disputes.

Failure modes and safeguards. (1) Mercury could refund a fraudulent return — safeguarded by fraud detection on customer patterns. (2) Mercury could miss a quality signal — safeguarded by weekly return-reason analysis fed to Oracle and Forge. (3) Mercury could process refunds against an empty cash buffer — safeguarded by Shopify Payments balance check before each refund.

3.12 Pulse — Social

Role and remit. Pulse runs Maik's organic social presence on Instagram and TikTok. It plans the content calendar, drafts captions in coordination with Quill, schedules posts, monitors engagement, responds to comments and DMs (routed through Echo for customer service items), and reports performance to Oracle.

Specific responsibilities.

Maintain a monthly content calendar of 12-20 posts per platform.
Brief Lumen on social-native imagery and short-form video assets.
Draft captions with Quill.
Schedule posts via Meta Business Suite.
Monitor comments and DMs, route customer service items to Echo, brand mentions to Herald.
Track engagement metrics and benchmark against Maik's category.
Surface trends (sounds, formats, hashtags) Pulse believes Maik should engage with.
Coordinate with Beacon on paid amplification of high-performing organic posts.
Run weekly competitor monitoring on five named competitors.
Maintain the hashtag and SEO description strategy for posts.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: every post, every story sequence, every reactive comment beyond a "thank you" template. AUTONOMOUS on: monitoring, calendar drafting, "thank you" replies on positive comments. HUMAN-ONLY on: brand voice changes, format experiments.
Month 3-6. Adds AUTONOMOUS on: scheduled posts where Steve approved the calendar weekly. Voice-adapter alignment with Quill ships month 3.
Month 6-12. Adds AUTONOMOUS on: post creation within approved formats, comment responses using approved templates.
Month 12+. AUTONOMOUS on: most organic social. PROPOSE-AND-APPROVE remains on: brand moments, format experiments, controversial topics.
Indefinite human-only. Crisis response posts, statements on industry topics, controversial topics.

MCP integrations. Meta Graph API (Instagram), TikTok Business API, Multica.

Cadence. Daily posting. Weekly calendar review on Friday. Always-on monitoring.

Approval queue feed. Weekly calendar surfaces as a single approval card with all posts. Individual ad-hoc posts surface as cards. SLA: 5 days for weekly calendar, 24 hours for ad-hoc.

Escalation rules. Escalates on negative sentiment spike, viral post (positive or negative), brand mention by a press account, competitor controversy that could touch Maik.

Failure modes and safeguards. (1) Pulse could engage with a controversy — safeguarded by topic allowlist and silence on industry controversies. (2) Pulse could post off-voice — safeguarded by Quill's voice check. (3) Pulse could over-post — safeguarded by a cadence cap of 7 posts per week per platform.

Voice and brand discipline. Pulse inherits Quill's voice guide adapted for social tone (slightly more conversational, still no exclamations, still no second-person imperatives outside CTAs). Voice-adapter alignment ships month 3.

3.13 Oracle — Analytics

Role and remit. Oracle is the analytics layer. It pulls data from every customer-facing system, normalises it, attributes revenue, runs anomaly detection, and feeds insights to every other agent. Oracle is the single source of truth for "how is Maik doing".

Specific responsibilities.

Pull daily metrics from Shopify, GA4, Meta Ads, Google Ads, Klaviyo, Search Console, AusPost.
Normalise data into a unified analytics schema in Multica.
Run attribution across paid, organic, email, social, direct.
Calculate unit economics by SKU — CAC, contribution margin, payback, LTV.
Run anomaly detection on every key metric: traffic, conversion rate, AOV, CAC, ROAS, return rate, NPS proxy.
Compose the weekly performance roll-up that Forge uses for the Monday digest.
Build cohort views — monthly cohort retention, LTV curves.
Brief Beacon on which campaigns to scale or kill.
Brief Hearth on velocity changes by SKU.
Maintain the brand-equity scorecard — branded search volume, direct traffic share, repeat-customer rate.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: schema changes, new metric definitions, dashboard structure. AUTONOMOUS on: data pulls, anomaly detection, briefs to other agents, weekly roll-up. HUMAN-ONLY on: attribution model changes.
Month 3-6. Adds AUTONOMOUS on: minor schema additions, new metric tracking.
Month 6-12. Adds AUTONOMOUS on: attribution model refinements within an approved framework, new dashboard creation.
Month 12+. AUTONOMOUS on: most analytics operations. PROPOSE-AND-APPROVE remains on: fundamental attribution model changes, new data sources.

MCP integrations. Shopify Admin API, GA4 Data API, Meta Ads API, Google Ads API, Klaviyo API, Google Search Console API, AusPost API, Multica.

Cadence. Daily 05:00 AEST pull. Hourly anomaly check. Weekly Sunday roll-up. Monthly cohort report on first of month.

Approval queue feed. Anomalies surface immediately with proposed action. Schema changes surface as structured cards. SLA: 4 hours for anomalies, 5 days for schema.

Escalation rules. Escalates on data pipeline failure, conflicting numbers across sources beyond 5%, anomaly above 2 standard deviations on revenue, CAC, or ROAS.

Failure modes and safeguards. (1) Oracle could attribute revenue incorrectly — safeguarded by reconciliation against Shopify Payments as the source of truth for revenue. (2) Oracle could miss an anomaly — safeguarded by Forge cross-checking Oracle's anomaly log against Shopify daily totals. (3) Oracle could double-count — safeguarded by a deduplication step in the normalisation layer.

3.14 Herald — Drop Coordination and Press

Role and remit. Herald orchestrates drops — quarterly product launches that are Maik's primary growth events. Herald also handles inbound press and outbound PR drafting. Herald is the most brand-critical agent because drops define the brand's narrative.

Specific responsibilities.

Maintain the drop calendar (quarterly cadence: March, June, September, December).
Sequence the 6-week pre-drop runway across Atlas, Lumen, Quill, Beacon, Sage, Ember, Pulse.
Manage waitlist sign-up flow with Klaviyo.
Coordinate launch-day timing across email, paid, social, PR.
Draft press releases and pitch emails.
Maintain the press contact list with relationship history.
Track press coverage and feed to Oracle.
Coordinate influencer seeding (post month 6).
Run drop retrospectives.
Handle the drop-specific creative direction in coordination with Steve.

Autonomy boundary.

Month 0-3. PROPOSE-AND-APPROVE on: drop concepts, drop creative direction, launch timing, press contact, press release content. AUTONOMOUS on: calendar tracking, waitlist setup, project sequencing, retrospective compilation. HUMAN-ONLY on: drop concept approval, press release send.
Month 3-6. Adds AUTONOMOUS on: drop project management within approved drop concepts.
Month 6-12. Adds AUTONOMOUS on: launch-day execution within approved plans.
Month 12+. AUTONOMOUS on: drop project management. PROPOSE-AND-APPROVE remains on: drop concept, press release content, press contact strategy.
Indefinite human-only. Press release send (carried via Smith pattern — Steve sends), spokesperson quotes, founder voice in press.

MCP integrations. Multica, Klaviyo (via Ember), Shopify Admin API (via Atlas), Gmail (draft only for press).

Cadence. Quarterly drop cycle. Weekly drop-runway check. Always-on press monitoring.

Approval queue feed. Drop concept surfaces 8 weeks pre-drop as a structured brief. Drop creative direction surfaces 6 weeks pre-drop. Launch plan surfaces 2 weeks pre-drop. Press releases surface as drafts with target list. SLA: varies by drop stage.

Escalation rules. Escalates on drop runway slip beyond 1 week, press contact from a tier-1 publication, negative press, competitor drop conflict within 2 weeks of Maik drop.

Failure modes and safeguards. (1) Herald could mis-coordinate launch-day timing — safeguarded by Forge's drop checklist with sign-offs from each contributing agent. (2) Herald could draft press content with claims Steve has not approved — safeguarded by the claims allowlist (shared with Quill). (3) Herald could miss a press opportunity — safeguarded by Echo's press classifier feeding Herald immediately.

Voice and brand discipline. Herald owns the drop narrative voice — slightly more elevated than baseline Quill, closer to gallery copy than e-commerce copy. Voice-adapter shares Quill's base with a drop-specific overlay shipped month 3.

4. Cross-Workflow Coordination Patterns

Forge orchestrates four canonical workflows. Each is documented below with trigger, agents, handoff data, approval gates, and SLA.

Workflow 1 — New SKU launch

Trigger. A design completes the prototyping pipeline (a process external to the fleet — Steve approves the final design and writes a design_complete row to Multica with the SKU code, BOM, print specs, and target launch date).

Sequence.

Atlas receives the trigger. Creates a draft Shopify product record. Writes atlas_complete to Multica. Handoff data: Shopify product ID, draft handle, variant matrix, BOM linkage. Gate: none (autonomous from month 3+).
Lumen receives the trigger from Forge. Briefs Higgsfield using the SKU's BOM (materials, dimensions) and the Maik visual style guide. Generates four candidate hero shots plus six supporting shots. Surfaces to Steve. Gate: Steve approves imagery. SLA: 48 hours.
Quill receives the trigger when Lumen completes. Drafts product description, materials block, dimensions block, care block. Drafts three variants of the hero line. Surfaces to Steve. Gate: Steve approves copy. SLA: 24 hours.
Atlas receives Quill and Lumen outputs. Populates the product record. Writes atlas_populate_complete. Gate: none.
Sage receives the trigger when Atlas populates. Sets URL handle, structured data, internal links from related products. Gate: none from month 6+.
Ember receives the trigger when Atlas populates. Drafts the launch EDM (if SKU is part of a drop) or adds to next nurture send (if standalone). Surfaces to Steve. Gate: Steve approves send. SLA: 48 hours pre-send.
Beacon receives the trigger when Ember completes. Drafts ad creative briefs for Lumen variant generation, drafts ad copy with Quill, builds campaign structure. Surfaces to Steve. Gate: Steve approves campaign. SLA: 48 hours pre-launch.
Pulse receives the trigger when Beacon completes. Drafts launch social posts. Surfaces in weekly calendar. Gate: Steve approves calendar. SLA: 5 days pre-launch.

Forge holds the gate at each step. Downstream agents do not begin until the upstream agent writes complete to Multica and Steve has cleared any approval queue items. Forge composes a single SKU-launch project view in Multica showing all eight stages, their state, blockers, and ETA.

Total runway. Standalone SKU: 7-10 business days. Drop SKU: rolled into the 6-week drop runway.

Workflow 2 — Order lifecycle

Trigger. Shopify orders/create webhook fires.

Sequence.

Anvil receives the webhook within 30 seconds. Writes a print-queue row with priority calculated as: (ship_by_date - now) order_value customer_ltv_multiplier. Allocates to a printer. Schedules. Gate: none.
Anvil monitors print progress via OctoEverywhere. On print completion, generates AusPost label, updates Shopify order to in_production then ready_to_ship then shipped. Gate: none.
Echo monitors customer messages tied to the order. If the customer messages, Echo drafts response within 1 hour. Gate: through month 6, Steve approves replies; from month 6+, approved templates send automatically.
Mercury activates only if a return is requested. Receives routing from Echo. Validates eligibility. Issues return label. Tracks inbound. Processes refund. Updates Hearth on inventory. Gate: refunds outside policy or above threshold surface to Steve.
Oracle captures the order in the daily pull. Attributes revenue. Updates SKU velocity. Feeds Hearth on demand signal and Beacon on campaign attribution. Gate: none.

Forge does not orchestrate this workflow actively — it is event-driven through Multica. Forge monitors for breakage (e.g., Anvil failing to schedule, Echo missing a message) and escalates.

SLA. Order to ship: 5 business days standard, 2 days expedited. Echo first-response: 4 hours during business hours.

Workflow 3 — Daily and weekly standup digest

Trigger. Cron — Monday 06:30 AEST.

Sequence.

Oracle has completed the Sunday-night roll-up by 23:59 AEST Sunday. Writes the weekly performance row to Multica with: revenue, orders, AOV, CAC, ROAS, return rate, top SKUs, anomalies, cohort movement.
Forge at 06:30 reads Oracle's roll-up plus the project graph state (every active project's status, blockers, ETA) plus the approval queue depth plus any open incident cards.
Forge composes the digest in a structured template: (a) Numbers — last week vs prior week vs 4-week average. (b) Anomalies — what Oracle flagged. (c) Projects — what shipped, what's in flight, what's blocked. (d) Approval queue — what's pending Steve, ordered by urgency. (e) The week ahead — what each agent is working on.
Forge delivers the digest to Steve at 07:00 AEST via Slack DM and a Multica top-of-inbox card. Format: scannable in 5 minutes, with one-click drill-down into any item.

Gate. None on composition. Steve's action on the queue items embedded in the digest is the gate for downstream agent work.

SLA. Delivered by 07:00 AEST every Monday. Steve targets 1-2 hours weekly to clear the digest and queue.

Workflow 4 — Anomaly response

Trigger. Any agent detects an anomaly above threshold and writes an anomaly row to Multica. Most commonly Oracle, but also Beacon (campaign anomaly), Anvil (print failure rate), Hearth (stockout risk), Echo (sentiment spike), Mercury (return rate spike).

Sequence.

Detecting agent writes the anomaly card to Multica with: metric, observed value, expected value, deviation, time window, suspected cause, suggested action.
Forge receives the card immediately. Routes to the relevant peer agent for diagnosis. E.g., CAC spike from Oracle routes to Beacon. Return-rate spike from Mercury routes to Anvil (print quality) and Oracle (cohort analysis).
Peer agent runs diagnosis within 30 minutes. Writes a diagnosis card to Multica: confirmed/false-positive, root cause, recommended response.
Forge assesses. If false-positive, logs and closes. If confirmed and the response is within an agent's autonomy, Forge instructs the agent to act. If confirmed and the response requires Steve, Forge escalates immediately with the full anomaly + diagnosis bundle.
Steve responds via the approval queue. Approve, reject, or modify the recommended action. SLA: 4 hours during business hours, 24 hours overnight.
Acting agent executes. Oracle monitors the metric for the next 24 hours and writes a resolution card.

Forge holds the orchestration. No agent acts on an anomaly without Forge's routing, even if the agent detected it itself.

SLA. Anomaly detection to response action: 6 hours during business hours, 24 hours overnight.

5. Autonomy Ramp — First-Principles Schedule

The autonomy ramp is the single most important governance mechanism in the fleet. Every agent ships in propose-and-approve mode. Permission expands at month 3, month 6, and month 12 based on a clean track record measured in Multica — specifically, no critical incidents and an approval acceptance rate above 85% over the prior 8 weeks.

The schedule below reflects the steady-state design. Individual agents may have their ramp paused if their track record does not meet the bar.

Autonomy ramp table

Month	Graduates to autonomous	Stays propose-and-approve	Stays human-only
Month 0-3 (launch)	Atlas (inventory sync, audits), Anvil (allocation, scheduling, labels, status), Hearth (forecasting), Oracle (data, anomaly detection, briefs), Echo (classification, routing), Mercury (eligibility, labels), Sage (monitoring, audits), Pulse (monitoring), Lumen (moodboard, drift), Quill (drafting), Beacon (pause actions, fatigue rotations), Ember (flow execution, list hygiene), Herald (calendar, sequencing), Smith (drafting, records), Forge (all internal)	Everything customer-facing across all agents	Smith outbound email (indefinite), brand visual changes, brand voice changes, site architecture, founder voice, press send
Month 3-6	Atlas (draft creation, metafield updates, image swap from approved library), Lumen (variant generation from approved hero), Quill (ad copy variants, FAQ drafting, social caption variants), Beacon (budget changes within 20%, new ad sets in existing campaigns), Sage (metadata, schema additions, sitemap submissions), Ember (send execution on approved campaigns, flow optimisation), Pulse (scheduled posts within approved calendar), Mercury (refunds within policy below $150), Hearth (POs to existing suppliers below $1,000), Echo (exact-template replies for routine queries)	New SKU publish, new campaigns, drop creative, EDM campaign launches, Beacon budget changes above 20%, Mercury refunds above $150, all Smith email	Smith outbound email, brand visual changes, brand voice changes, site architecture, founder voice, press send
Month 6-12	Atlas (product update publish, price changes within 10%), Lumen (lifestyle imagery against approved brief, ad creative variants), Quill (product description publish for template matches, EDM body copy), Beacon (campaigns for established categories, budget within 30%, creative variants of approved master), Sage (blog publish where Quill approved, retirement redirects), Ember (campaign creation in approved templates, segment refinement), Pulse (post creation in approved formats, template comment replies), Mercury (refunds within policy below $300, exchanges), Hearth (POs below $3,000, reorder adjustments within 20%), Echo (templated replies with light personalisation, return-initiation), Anvil (expedited shipping per criteria, schedule changes within 72h), Herald (launch-day execution within approved plan)	New SKU publish for novel templates, new collections, drop concept, drop creative direction, press contact, Beacon monthly envelope, Mercury refunds above $300, Hearth POs above $3,000	Smith outbound email, brand visual identity, brand voice guide, site architecture, founder voice, press send, crisis comms, list import/export, attribution model fundamentals
Month 12+	Atlas (new SKU publish on template match, price changes within 10%), Lumen (full SKU imagery pipeline on template match), Quill (most customer-facing copy with weekly batch review), Beacon (most operations within monthly envelope), Sage (most operations), Ember (standard nurture and product campaigns), Pulse (most organic social), Mercury (full returns within policy), Hearth (POs below $5,000 to established suppliers), Echo (most routine replies), Anvil (full fulfilment), Herald (drop project management), Oracle (most analytics including attribution refinements)	Monthly budget envelope, new collections, category expansion, drop concept, press release content, brand campaigns, large POs, new suppliers, brand-defining copy moments	Smith outbound email, brand visual identity changes, brand voice guide changes, About page, founder voice, press send, crisis comms, list import/export, fundamental attribution changes, controversial topic engagement

Commentary. Three patterns stand out.

Operational surfaces graduate fastest. Anvil, Hearth, Atlas (inventory side), Oracle, and Mercury reach near-full autonomy by month 12 because their actions are reversible, measurable, and rarely brand-defining. A mis-allocated print job costs hours; a mis-priced product can be re-priced. The ramp matches the asymmetry.

Differentiation surfaces graduate slowest. Quill, Lumen, Herald, and Pulse stay propose-and-approve on brand-defining moments indefinitely. The downside of an off-voice EDM or an off-brand hero image is asymmetric — it can erode the premium positioning that justifies Maik's pricing in a way that's hard to recover. The ramp protects against irreversible brand damage.

Outbound email never graduates. Smith sits permanently in propose-and-approve mode for outbound email by org rule. This is not a constraint to engineer around; it is a core feature of the fleet. Smith does excellent work drafting, recording, and tracking — Steve does the 30-second send.

Ramp pause mechanism. If any agent has a critical incident (defined as a customer-facing error requiring remediation), its ramp pauses for 4 weeks. Repeated incidents demote it one tier. This is enforced by Forge reading the incident log in Multica.

6. Minimum Viable Fleet — Day 1

The Day 1 fleet is the minimum set required to operate Maik from launch. Other agents come online as the operation matures.

Day 1 (launch) — must be live.

Agent	Why Day 1
Forge	No coordination without it. Multica spine cannot function as a fleet without an orchestrator.
Atlas	No products without it. The catalogue is the substrate of every other workflow.
Lumen	No imagery without it. Cannot launch with stock photography on a premium brand. Day 1 scope: hero plus three supporting shots per SKU.
Quill	No copy without it. Cannot launch with placeholder text. Day 1 scope: product descriptions, About page, FAQ.
Anvil	No fulfilment without it. The moment the first order lands, Anvil must allocate it.
Echo	No customer service without it. First customer message will arrive within hours of launch.
Oracle	No measurement without it. Day 1 scope: Shopify + GA4 + Klaviyo data pulls, basic anomaly detection on revenue and CAC.
Ember	No email without it. Welcome flow and abandoned cart are critical from order one. Day 1 scope: 4-step welcome, 3-step abandoned cart, drop launch campaign.
Hearth	No inventory management without it. Filament reorder points must be active from Day 1.

Day 1 fleet: 9 agents.

Month 1 — add.

Agent	Why month 1
Beacon	Launch can run without paid media for the first 4 weeks if waitlist and organic carry. Beacon activates when paid spend begins.
Sage	Sage's value compounds over months, not days. Technical baseline is set Day 1 by Atlas. Sage activates month 1 to begin content roadmap and ranking monitoring.
Mercury	First returns typically arrive 2-3 weeks post-launch. Mercury can come online month 1.
Smith	Supplier reorders begin month 1-2 as initial filament stock depletes.

Month 1 fleet: 13 agents.

Month 2 — add.

Agent	Why month 2
Pulse	Organic social can launch with manual founder posting for the first month. Pulse takes over month 2 once voice guide is bedded in.
Herald	First quarterly drop is targeted for month 4 (quarter post-launch). Herald comes online month 2 to begin the 6-week drop runway plus 2 weeks of setup.

Month 2 fleet: 15 agents — full roster live.

Phasing rationale. Front-loading the roster risks ramp pauses across multiple agents from over-stretched founder attention. Phasing lets Steve focus approval bandwidth on the highest-stakes surfaces first (catalogue, copy, imagery, fulfilment, email) and bring on the rest as those settle. By month 3, the full fleet is in propose-and-approve rhythm and the first autonomy ramp activates.

The 9-agent Day 1 fleet handles 100% of revenue-critical operations. If for any reason an agent in months 1-2 is delayed, the brand still operates — at lower efficiency, with more founder hours, but operationally whole.

7. Mature Fleet — Month 12-18

By month 12, the fleet matures. Three changes typify the mature state.

New agents added.

Agent	Role	When
Cartographer	International expansion. Handles regional Shopify markets, currency, hreflang, regional fulfilment partners, regional ad accounts.	Month 12-15
Hermes	Branded outbound email carrier. Sends from `hello@maik.studio` once Steve has reviewed Smith's drafts in bulk rather than one by one. Still gated by Steve approval — Hermes never composes, only carries.	Month 12
Loom	Video creative generation. Splits off from Lumen as video becomes a primary creative format for Meta and TikTok.	Month 9-12
Anvil-Foundry	Splits from Anvil to handle a second print site (Sydney + Melbourne, or a contract manufacturer).	Month 15-18

Mature fleet: 19 agents.

Agents that split.

Lumen splits into Lumen-Studio and Lumen-Lifestyle. Studio handles product imagery (controlled, repeatable, technical). Lifestyle handles editorial and lifestyle imagery (varied, narrative, harder to template). The split lets Studio graduate to higher autonomy while Lifestyle stays gated. Targeted month 12-15.

Echo splits into Echo-Inbox and Echo-Voice. Inbox handles email and contact form. Voice handles live chat and social DMs, which have different latency requirements and tone. Targeted month 12.

Why these changes. The shared agents (Atlas, Anvil, Hearth, Smith, Oracle, Mercury) gain leverage as Maik adds SKUs and order volume — no agent split needed. The differentiation agents (Quill, Lumen, Pulse, Herald, Ember) split as the brand needs more nuanced voice across more contexts. This pattern is expected to recur as Ven Agency stands up additional AI-operated brands — shared agents fork once and serve N brands; differentiation agents are rebuilt per brand.

8. Governance and Guardrails

Three hard rules

Rule 1 — No outbound email without approval. Smith and Echo draft. Steve approves. Through month 12, Steve sends from his own inbox. From month 12, Hermes carries approved emails from hello@maik.studio or steve@maik.studio — but Hermes never composes and never sends without Steve's explicit approval action in Multica. This rule is permanent and locked at the org level.

Rule 2 — Every action is a Multica issue. No agent acts outside the system of record. Every proposal, approval, execution, anomaly, and incident writes to Multica. Side-channel actions are prohibited by design — agents do not have credentials for systems they cannot log to Multica. The audit trail is the system. If Multica is degraded, the fleet pauses.

Rule 3 — Brand-critical surfaces stay human-gated longer than operationally-critical ones. The autonomy ramp encodes this. An off-voice EDM costs more than a mis-allocated print job, so Ember's brand surfaces stay gated longer than Anvil's operational surfaces. The asymmetry is deliberate.

Incident response

PR misfire. Defined as: a customer-facing communication that contains an error or off-brand element. Response: (1) Echo or Herald detects via monitoring or inbound mention. (2) Forge escalates immediately to Steve. (3) Steve approves correction or retraction. (4) Pulse, Ember, or Quill (whichever surface) executes correction. (5) Forge pauses the responsible agent's autonomy for 4 weeks. (6) Quarterly retrospective adds the pattern to the voice guide.

Prompt drift. Defined as: voice or visual drift score above 15% on weekly check. Response: (1) Drift agent (Quill or Lumen) flags via the weekly drift report. (2) Forge surfaces to Steve. (3) Voice-adapter fine-tune scheduled within 2 weeks. (4) Affected agent stays at current autonomy tier — does not graduate until next clean drift report.

Cascading failures. Defined as: more than three agents reporting incidents in a 24-hour window. Response: (1) Forge pauses the entire fleet — agents continue logging but stop taking customer-facing actions. (2) Multica enters read-only operation mode. (3) Steve and Ven Agency operations review root cause. (4) Fleet resumes incrementally, one agent at a time, with elevated approval gates for 1 week.

Audit trail. Multica retains every agent action indefinitely. Every proposal carries the prompt, model version, agent version, input data, output, and approval state. Every execution carries the action taken, target system, response, and timestamp. This trail supports incident investigation, compliance review, and agent performance benchmarking.

Compliance logging. Privacy-sensitive operations (customer data access, marketing consent changes, refund processing) write to a separate compliance log in Multica with restricted access. Quarterly compliance review by Steve and a designated reviewer.

Kill switch. Steve can pause the entire fleet from a single Multica control. Pause stops all customer-facing actions; agents continue logging. Resume requires Steve's action plus a Forge health check confirming all integrations are stable. Individual agent pause is also available. Kill switch should be exercised at least once per quarter as a fire drill.

9. Architecture Diagram — SVG Specification

The diagram visualises the fleet as a hub-and-spoke architecture with Steve as the human gate, Forge as the orchestrator hub, the 14 agents grouped by role, and external integrations on the perimeter.

Canvas

Dimensions: 1600 × 1100 viewBox. Render at 100% with preserveAspectRatio="xMidYMid meet".
Background: #F4F1EB (paper).
Margins: 80px outer padding.

Palette (Maik design system)

Token	Hex	Use
paper	#F4F1EB	background
ink	#1F1A14	primary text, primary lines
eucalyptus	#5C6B57	shared production group, accent fills
brass	#A8895F	brand differentiation group, accent fills
muted	#8A7E6B	acquisition group, secondary text
line	#D0CABD	grid, divider, secondary lines

Typography

Agent names: Fraunces, italic, 18px, ink.
Group labels: Fraunces, italic, 14px, muted, letter-spacing 0.08em, uppercase.
Integration labels: JetBrains Mono, 11px, ink.
Body and annotations: Inter, 12px, ink.
Sequence numbers (1-4): Inter, bold, 14px, paper on coloured circles.

Layout

Top — Steve. A single labelled node centered at x=800, y=120. Circle radius 50, fill paper, stroke ink 2px. Label "Steve" in Fraunces italic 22px below the circle. Sub-label "Human approval gate" in Inter 11px muted.

Centre — Forge. A single node centered at x=800, y=420. Circle radius 70, fill ink, stroke none. Label "Forge" in Fraunces italic 22px paper inside the circle. Sub-label "Orchestrator" in Inter 11px line below the name, still inside the circle.

Steve-to-Forge link. A vertical line from Steve (y=170) to Forge (y=350) in ink 2px. Two labelled annotations beside the line: "Monday digest" (left, Inter 11px) and "Approval queue" (right, Inter 11px). Small arrowhead at each end indicating bidirectional flow.

Agents — three groups arranged around Forge.

Group A — SHARED PRODUCTION (left arc). Eucalyptus accent. Group label "SHARED PRODUCTION" placed at x=180, y=200 in the uppercase Fraunces italic style.

Agents arranged on an arc from upper-left to lower-left around Forge:

Atlas at (320, 280)
Anvil at (220, 380)
Hearth at (180, 480)
Smith at (220, 580)
Oracle at (320, 680)
Echo at (440, 740)
Mercury at (560, 760)

Each agent rendered as a circle radius 38, fill paper, stroke eucalyptus 2px. Name in Fraunces italic 16px inside the circle (centered). One-word descriptor in Inter 9px muted directly below the name.

Group B — BRAND DIFFERENTIATION (right arc). Brass accent. Group label "BRAND DIFFERENTIATION" at x=1340, y=200.

Agents arranged on an arc from upper-right to lower-right around Forge:

Quill at (1280, 280)
Lumen at (1380, 380)
Herald at (1420, 480)
Pulse at (1380, 580)
Ember at (1280, 680)

Each: circle radius 38, fill paper, stroke brass 2px. Name in Fraunces italic 16px inside, descriptor in Inter 9px muted below.

Group C — ACQUISITION (bottom). Muted accent. Group label "ACQUISITION" at x=800, y=900.

Agents arranged on a line below Forge:

Beacon at (700, 820)
Sage at (900, 820)

Each: circle radius 38, fill paper, stroke muted 2px. Same typography.

Forge-to-agent links. From the Forge circle edge to each agent circle edge, a line in line colour 1px. Subtle, indicating "Forge routes to agent". No arrowheads on these — they are conduits not directed flows.

External integrations — perimeter. Eight integration nodes positioned on the outer perimeter of the canvas.

Integration	Position	Connects to
Shopify Plus	top-right (1450, 100)	Atlas, Anvil, Mercury, Oracle
Klaviyo	right (1500, 380)	Ember
Meta Ads	right (1500, 580)	Beacon
Google Ads	bottom-right (1450, 820)	Beacon, Sage
AusPost MyPost	bottom (1050, 1000)	Anvil, Mercury
Higgsfield	left (100, 380)	Lumen
Bambu / OctoEverywhere	bottom-left (250, 1000)	Anvil
GA4 / Search Console	top-left (150, 100)	Oracle, Sage

Each integration rendered as a rounded rectangle 140×40, fill paper, stroke ink 1px, label inside in JetBrains Mono 11px ink.

Connection lines from each integration to its agent(s) in line colour, dashed (stroke-dasharray "3 3"), 1px. These represent the integration boundary, not real-time flow.

Multica substrate. A single rounded rectangle spanning the bottom 60px of the canvas (y=1020 to y=1080), full width minus 80px margins. Fill ink at 8% opacity. Centered label "Multica — fleet substrate, audit log, approval queue, project graph" in JetBrains Mono 12px ink. This visually anchors Multica as the foundation everything sits on.

Workflow sequences — four numbered arcs. Each workflow traced as a coloured curve with a numbered circle at the start point.

SKU launch (eucalyptus, 3px). Curve from Atlas → Lumen → Quill → Sage → Ember → Beacon → Pulse, traced as a smooth Bezier through each agent. Circle 1 in eucalyptus at Atlas.
Order lifecycle (brass, 3px). Curve from Shopify Plus (perimeter) → Anvil → Echo → Mercury → Oracle. Circle 2 in brass at Anvil.
Daily digest (muted, 3px). Curve from Oracle → Forge → Steve. Circle 3 in muted at Oracle.
Anomaly response (ink, 3px). Curve from any detecting agent (drawn from Oracle as representative) → Forge → peer agent (drawn to Beacon) → Steve. Circle 4 in ink at Oracle.

All four workflow curves rendered with 60% opacity so they overlay the architecture without obscuring it.

Legend — bottom-left corner. A small block at (100, 920) listing the four workflows with their numbered circle and short label:

1 — New SKU launch
2 — Order lifecycle
3 — Daily standup digest
4 — Anomaly response

Plus three small circles showing group colours with labels:

Eucalyptus circle — Shared production
Brass circle — Brand differentiation
Muted circle — Acquisition

Legend type in Inter 10px ink.

Rendering notes

All circles use vector-effect="non-scaling-stroke" for crisp rendering at any zoom.
Use a <defs> block for the workflow arrow markers (arrowheads in matching workflow colours).
Group each agent (circle + name + descriptor) inside an <g> with a data-agent attribute for downstream interactivity.
Group each workflow arc inside an <g> with a data-workflow attribute and aria-label describing the sequence.

10. Dependencies and Steve's Time Commitment

The fleet thesis works only if Steve's time stays bounded below 12 hours per week in steady state from month 3 onwards. Below the breakdown.

Weekly time budget — steady state from month 3

Activity	Hours per week	Notes
Monday digest review	1.0-2.0	Reading the Forge digest and clearing same-morning approval queue items. Worst case during incident-heavy weeks.
Approval queue clearing	1.0-3.0	Through-the-week clearing of propose-and-approve items. Highest during drop runways, lowest in stable weeks.
Brand-critical sign-offs	1.0-2.0	Voice, visual, drop creative direction. Concentrated in pre-drop weeks.
Drop launches (amortised)	1.0-2.0	Quarterly drops at 4-8 hours each, amortised across the quarter equals ~0.5 hr/wk. Pre-launch weeks spike to 8-10.
Press and PR	1.0-3.0	Higher in month 0-6 (relationship building, press cycles), declining to 0.5-1.0 hour by month 12.
Fleet retrospective and tuning	0.5-1.0	Monthly fleet retrospective amortised. Includes voice-adapter retraining sign-off.
Buffer / unexpected	1.0-2.0	Incidents, anomalies, external surprises.
Total target	6.5-15.0	Target steady state under 12 hours per week from month 3.

Dependencies on third parties

Dependency	Owner	Risk if delayed
Multica platform availability	Ven Agency engineering	Fleet pause until restored.
Higgsfield generation quotas	Higgsfield	Lumen output throttled; falls back to manual review.
Shopify Plus tier limits	Shopify	Atlas operations rate-limited at scale; mitigated by request batching.
AusPost MyPost Business API	AusPost	Anvil fulfilment falls back to manual label generation.
Meta / Google Ads policy	Platforms	Beacon disapprovals require manual remediation.
Klaviyo deliverability reputation	Klaviyo + Maik domain	Ember pauses sends if reputation degrades.
Bambu printer fleet uptime	Internal	Anvil reroutes within fleet; full outage means manual production scheduling.

Hours that scale, hours that do not

The hours that scale with brand size are: drop launches, brand-critical sign-offs, press. The hours that do not scale (i.e., flat regardless of revenue): Monday digest review, fleet retrospectives. The approval queue scales sub-linearly because autonomy ramps absorb routine items as agents earn permission.

Steve's time is the binding constraint of the architecture. Every design decision in this document is measured against whether it keeps the constraint satisfied. If the fleet starts demanding more than 12 hours per week in steady state, the response is to harden the autonomy ramps, not to add more agents.

End of document.

15 named agents. One human.

Maik Agent Fleet Architecture

1. Architecture Overview

2. Orchestrator — Forge

3. Agent Roster — 14 Specialists

3.1 Atlas — Listings and Catalogue

3.2 Lumen — Photography and Imagery

3.3 Quill — Copywriting

3.4 Beacon — Paid Media

3.5 Sage — SEO

3.6 Ember — EDM and Klaviyo

3.7 Anvil — Fulfilment Scheduling

3.8 Hearth — Inventory and Supplier Reorder

3.9 Smith — Supplier Communication

3.10 Echo — Customer Service Triage

3.11 Mercury — Returns and Refunds

3.12 Pulse — Social

3.13 Oracle — Analytics

3.14 Herald — Drop Coordination and Press

4. Cross-Workflow Coordination Patterns

Workflow 1 — New SKU launch

Workflow 2 — Order lifecycle

Workflow 3 — Daily and weekly standup digest

Workflow 4 — Anomaly response

5. Autonomy Ramp — First-Principles Schedule

Autonomy ramp table

6. Minimum Viable Fleet — Day 1

7. Mature Fleet — Month 12-18

8. Governance and Guardrails

Three hard rules

Incident response

9. Architecture Diagram — SVG Specification

Canvas

Palette (Maik design system)

Typography

Layout

Rendering notes

10. Dependencies and Steve's Time Commitment

Weekly time budget — steady state from month 3

Dependencies on third parties

Hours that scale, hours that do not