Agency Brains Design.02
The companion to № 01. Thirty turns from an agency with one implicit, undeclared brain to an agency running a typed cognition stack under a Brain Council. Each turn is one move. Each move cites the AUX patterns, principles, heuristics, and trust stages it operationalises. Read it as a route. Use it as a checklist. Hand it to a client as the engagement made visible.
№ 01 named the defect — agencies are shipping stochastic answers to deterministic questions because nobody declared the brains. This document is the route from that state to a working cognition stack. Thirty turns, five legs, four trust stages unlocked along the way.
Every turn cites the AUX canon. Trust Canvas when the canvas is the working artifact. H06 when a heuristic is being satisfied. Pattern P05 when a brain operationalises Memory in Motion. Stage 02 when Contextual Trust becomes available. This is deliberate — it is what makes the work auditable, defensible, and forwardable. If a turn cannot cite the canon, it is not in the route.
Six months from Turn 01 to Turn 30. Two principals in the room, one Brain Council standing up around them, three brains shipped under a router that the agency owns at handoff. That is the destination.
How to read a turn.
Every turn in the route shares the same shape. Number, action verb, body, three pieces of metadata. The metadata is what makes the turn auditable and the engagement defensible.
What you touch.
The working artifact for this turn — Trust Canvas, Brain Manifest, AI Surface Register, Decision Tree, Audit Log Schema. If a turn doesn't produce or update an artifact, it isn't in the route.
What it satisfies.
The patterns (P01–P06), principles (1–8), heuristics (H01–H10), and trust stages (01–04) this turn operationalises. Every move ties back to the published canon. Forkable. Citable.
How you know it's done.
The observable signal that this turn is complete and the route can proceed. Falsifiable, not aspirational. If you can't point to the signal, you haven't completed the turn.
Before Turn 01.
The route starts assuming three things are already true. If they aren't, the engagement begins with a Turn 00 — a one-week scoping conversation to make them true. Most agencies need this. It is not a failure; it is a normal first week.
Reference materials
- AUX canon — Manifesto, six patterns, eight principles, ten heuristics, four trust stages. auxfirst.com/news/what-is-agentic-user-experience
- AUX Heuristics — the ten reimagined for agents. auxfirst.com/heuristics
- TrustKit — schemas, agent-spec, trust-contract, trust-gap-taxonomy. github.com/auxfirst/trustkit
- Trust Canvas — printed and unprinted copies. Figma Community
- № 01 — Agency Brains Design — the conceptual spec the atlas implements
- MCP for Advertising — the plumbing layer this routes through
Stakeholders in the room
- Sponsor — Chief Digital Officer or equivalent. Owns the engagement budget and the political cover.
- Engineering lead — CTO or VP Engineering. Owns the build phase and the MCP layer this rides on.
- Strategy lead — head of planning, head of strategy, or chief strategy officer. Owns Brief Brain and Account Brain.
- Brand lead — owner of brand book, claim policy, voice rules. Owns Brand Brain.
- Effectiveness lead — head of measurement, head of effectiveness, or analytics lead. Owns Effectiveness Brain.
- Legal / compliance — observer in Leg 01, active in Leg 02 and Leg 04. Owns Compliance Brain.
The Brain Council is named in Leg 02 — but the people who will sit on it need to be in the room from Leg 01. The names should not be a surprise to anyone by Turn 06.
See the agency as it actually reasons.
Visibility of Agent Intent & Action — the agent should always make its intent legible before or during action. AUX Heuristic H01
You can't fix what you can't see. Leg 01 makes the agency's current implicit cognition visible — every AI surface, every shadow surface, every implicit reasoning type. The leg ends with a leadership readout the C-suite can act on.
Openthe Trust Canvas with the stakeholder council.
Ninety minutes, the named six stakeholders, a printed Trust Canvas on the wall. Do not talk about brains yet. Walk the canvas as it is meant to be walked — autonomy boundaries, handoffs, decision points — and have the room name the AI surfaces it knows about. List, don't judge. The first artefact of the engagement is a marked-up canvas the room has touched.
Mapevery AI surface and reasoning call — including the shadow ones.
Outside the boardroom, walk the floors. Strategists, traders, account leads, creative directors, finance. Find every place an LLM is being called — sanctioned or unsanctioned. Most agencies discover three to five times more shadow surfaces than the leadership team named in Turn 01. Document each: surface, owner, question types, model called, implicit reasoning type.
Classifythe implicit reasoning type on every surface.
Per surface, ask the same question: what kind of reasoning is this AI being asked to perform, and what kind is it actually performing? In nearly every case the gap is the same — questions need rule-based, causal, or probabilistic answers; the system is delivering stochastic-generative. Tag every gap with a trust-gap classification from the TrustKit taxonomy.
Runthe ten heuristics against the top three high-traffic surfaces.
Pick the three most-used AI surfaces from the register. Run all ten AUX heuristics against each — H01 through H10, in order, written down. Every agency we have ever audited fails at least four of the ten on at least one surface. Failures cluster on H01, H05, H06, H09. This is the data that puts the readout on solid ground.
Deliverthe Defect Inventory and readout to the C-suite.
Single session, 60–90 minutes, sponsor in the room. The deliverable is a Defect Inventory — Brain Map, Surface Register, Reasoning Mismatch Table, three Heuristic Scorecards, trust-gap classifications, prioritised risk list. The readout closes Phase 0 and authorises Phases 1–4. Anything the room cannot defend in this readout doesn't pass into the build.
Declare the brains. End shadow cognition.
Clarify Before Commit — before any high-stakes action, the agent pauses and confirms intent. Intelligent friction protects time, reputation, and trust. AUX Principle 3
Leg 02 turns the implicit into the declared. The agency authors its brain taxonomy, names a human owner for each brain, drafts its first manifest, and pressure-tests the manifest against the AUX principles before any code is written.
Convenethe Brain Council pre-meeting.
The six stakeholders from Pre-flight + two engineers + the auxfirst principal. Two-hour working session. The agenda is set: walk the 14-type taxonomy, declare the ten brains, assign owners, schedule the manifest reviews. The room agrees on cadence and on the rule that no brain ships without an owner and a manifest.
Walkthe fourteen-type reasoning taxonomy.
The council reads through the taxonomy from № 01 — deterministic, rule-based, probabilistic, statistical, Bayesian, causal, heuristic, fuzzy, stochastic, generative, adaptive, agent-based, chaotic, emergent — and labels every Defect Inventory entry with the reasoning type it needed. The exercise is short. The vocabulary is what matters.
Declarethe ten brains this agency will run.
Council names the brain library. Default starter: Brand, Compliance, Brief, Effectiveness, Forecast, Creative (the six core) plus Audience, Account, Market, Bidder (the four extensions). Adjust per the agency's mix — a creative shop may add Editorial Brain; a trading desk may add Supply-Path Brain. The number is fixed at ten in the first phase to avoid sprawl.
Assigna named human owner to every brain.
One brain, one human, one accountability line. Brand Brain to the brand lead. Compliance Brain to legal. Brief Brain to head of strategy. Effectiveness Brain to head of measurement. No anonymous brains. The owner is responsible for the manifest, the knowledge sources, and the trust signature. They sit on the Brain Council. They sign off changes to their brain.
Draftthe Brand Brain manifest as the warm-up.
Brand is the simplest manifest to write — deterministic, rule-based, citable. Owner drafts, auxfirst principal facilitates. Five required fields: reasoning type, knowledge sources, output contract, trust signature, router tags. The format is published as brain-spec.schema.yaml — TrustKit-aligned, sitting next to agent-spec.schema.yaml and trust-contract.yaml in the repo. The first manifest sets the precedent for the next nine.
Pressure-testthe manifest against the eight AUX principles.
Council session. Run the Brand Brain manifest through each of the eight principles in turn. Does the output contract taper transparency (P1)? Does the brain prompt back when inputs are missing (P2)? Does it clarify before committing (P3)? Will it push back on a brand-violating draft (P4)? The exercise is short and decisive. Manifest is amended or signed off. The same exercise will run for every subsequent brain manifest.
Build the layer that chooses.
Consistency of Behavior, Not Interface — the interface can adapt, but decision logic, trust boundaries, and tone must stay stable. AUX Heuristic H10
The meta-brain is the agency's consistency guarantee. Surfaces can change shape; the router does not. Leg 03 designs it — five signals, a rule-based decision tree, fallback policy, audit log, governance gate.
Definethe five routing signals.
Reproducibility required? Audit / disclosure required? Causal claim being made? Forecast horizon involved? Variation the goal? Each signal is binary or scaled, declared in code, version-controlled. The router classifies on these five — no others. Adding a sixth signal is a council decision; never a developer decision.
Drawthe rule-based decision tree.
The tree maps signal combinations to brain invocations. It is rendered as a one-page diagram — readable by non-engineers, defensible by engineers, version-controlled by both. The router is rule-based by mandate. An LLM "deciding" which brain to call is the very stochastic-routing problem the architecture exists to end.
Writethe fallback policy and the shadow-cognition flag.
When the rules can't resolve, the router does not guess. It routes to the declared default (typically Brief Brain), and emits a shadow-cognition flag for human review. Over the first ninety days these flags are the most valuable signal in the entire system — they're where the next brains in the library are born.
Buildthe audit log schema.
Every routed request writes a record: brain ID, reasoning type used, knowledge sources read, trust signature returned, latency, signals matched. This is the audit substrate. It is what the CTO presents to the holdco risk committee, what the CFO uses to defend AI spend, and what the agency uses to answer "who said this".
Approvethe routing policy as the only allowed entry point.
The sponsor signs a one-page Routing Policy — no production-facing AI surface may bypass the router. The policy is binding inside the agency and tracked. Surfaces that currently bypass are listed with sunset dates. Phase 2 closes with the rule that the router is the new front door.
Three brains, live, under the router.
Graceful Handling of Uncertainty — when unsure, the agent asks, escalates, or returns a partial result, never hallucinates. Failure is a first-class design surface. AUX Heuristic H06
Brand first, Brief second, Effectiveness third. Each brain ships under the router, satisfies the AUX heuristics declared in its manifest, and unlocks the next trust stage. Eight to twelve weeks. The first time the agency sees its own work change shape.
Wirethe Brand knowledge MCP servers.
Three MCP servers, read-only, scoped: brand book, claims policy, asset rights register. Built per the MCP for Advertising piece. The plumbing is the easy part — the corpus curation is where the real work lives. Brand lead curates each corpus before connection.
Implementthe Brand Brain reasoning engine.
Rule engine (OPA, Drools, or custom) consuming the MCP corpora. LLM is used for input classification only, never adjudication. Every output is structured: verdict, citations, jurisdiction, rule version. The output contract from the manifest is the test — if the engine doesn't honour it, the engine isn't done.
ConnectBrand Brain to the router. Ship to the briefing tool.
The router invokes Brand Brain on every request that touches client-facing or brand-position output. The briefing tool is the first surface to integrate. Strategists see, for the first time, typed brand verdicts with rule citations inline in their workflow. The audit log starts writing. The trust dividend begins.
Buildthe Brief corpus + similarity engine.
Historical briefs ingested, tagged by client, sector, brief type, outcome. Embedding model + retrieval layer. This is where Memory in Motion (P05) becomes a real surface in the agency. Strategists' past judgement, encoded.
ImplementBrief Brain — heuristic + fuzzy logic.
Brief Brain takes a brief, returns: similarity to historicals, structural completeness flag, missing-field list, suggested interpretation. Fuzzy match scores are exposed, never hidden. The strategist sees what the brain is uncertain about and can intervene — exactly the H06 contract.
ShipBrief Brain to the strategist surface.
Every brief that enters the agency now passes through Brief Brain on the way to a strategist. Suggested similar historicals, completeness flag, interpretation map. This is the turn strategists feel the engagement daily. Stage 02 Contextual Trust becomes a target the agency is actually approaching.
ConnectEffectiveness Brain to its data sources.
MMM outputs, holdout/control test data, lift studies, attribution model artifacts. The Effectiveness Brain is only as good as the data it cites. Effectiveness lead owns the curation; legal reviews the data-use boundaries. Causal-clean datasets only.
Implementcausal inference engine + publish the methodology card.
Causal engine — DoWhy, EconML, or custom — wired to the data sources. Every response returns: lift estimate, counterfactual, confidence band, methodology card, assumption log. The methodology card is the killer artefact — it's what makes the agency's effectiveness claims defensible against any client procurement team that asks "how did you arrive at that".
ShipEffectiveness Brain to the pitch + capability deck workflow.
Every causal claim in every capability deck, case study, and pitch document is generated by Effectiveness Brain or marked as unverifiable. This is the trust-architecture turn that the engagement was built for. The agency's effectiveness story acquires a substrate that almost no competitor has. Stage 03 Judgment Trust becomes the visible direction of travel.
Hand the stack to the Brain Council.
Trust is Dynamic, Not Static — agent behaviour must evolve as trust develops. New users get cautious, gated behaviour; experienced users get faster, more autonomous behaviour. AUX Heuristic H04
The engagement does not end with three brains shipped. It ends with the agency owning the cognition stack and able to extend it. Leg 05 stands up the operating model, instruments the telemetry, drafts the roadmap, and hands the keys over.
Stand upthe Brain Council with a written charter.
The council becomes a standing body. Charter declares: membership, cadence (monthly working session, quarterly full review), authorities (manifest approval, brain additions, shadow-flag adjudication), reporting line (sponsor). The council is the most important organisational artefact this engagement leaves behind.
Establishthe quarterly brain review cycle.
Every brain reviewed quarterly — manifest, output samples, accuracy drift, owner sign-off. A brain that misses two consecutive quarterly reviews is suspended from the router. The discipline is what keeps the stack alive past month six, when most consulting engagements quietly decay.
Instrumentthe trust-gap telemetry.
Shadow-flag rate per brain, override rate per brain, output-contract violation rate, latency, audit-log completeness. Dashboards built. The dashboards become the council's standing agenda — they say which brains are quietly drifting and which are quietly outperforming, before any client notices either.
Draftthe Library Roadmap — the next three brains.
Council selects the next three brains from the agency's remaining seven (Compliance, Audience, Account, Forecast, Creative, Market, Bidder). Each gets a target quarter and an owner. This is what makes the engagement self-extending — the agency does not need us back to build brains four through ten.
Hand off— auxfirst transitions to advisor.
Closing ceremony. The Brain Council formally takes ownership: manifests, schemas, dashboards, router, charter, roadmap. auxfirst transitions to a quarterly advisor cadence — present at reviews, available for the next-brain consultations, called on for trust-gap escalations. The engagement is not over; the dependency is.
What it feels like when you've arrived.
A strategist opens a brief on Monday. The briefing tool returns a typed similarity panel from Brief Brain, a brand-rule pass from Brand Brain, and a flag from Compliance Brain for a regulated-category claim. None of those answers came from asking the AI. They came from three declared brains, each with a manifest, each with an owner.
On Thursday the agency pitches. Three case studies in the deck carry methodology cards from Effectiveness Brain. The procurement team on the client side has never seen a deck like that before. They notice. They remember.
On Friday afternoon the Brain Council meets. Dashboards open. Audience Brain is showing increased shadow-flag rate on healthcare cohorts. The council discusses, decides to update the manifest, schedules the change for Q+1. Nothing is improvised. Everything is logged. The agency is operating cognition.
What the trust architecture looks like, six months in
Functional Trust
Achieved on Brand and Compliance surfaces. Reliable, citable, reproducible.
Contextual Trust
Achieved on Brief and strategist surfaces. Memory in Motion is daily.
Judgment Trust
In sight on pitch surfaces via Effectiveness Brain methodology cards.
Advocacy Trust
Conditions in place. The next 12 months under the Brain Council.
The agency stops asking "what did the AI say". It starts asking "which brain answered, on what basis, with what data, under what method". That single question shift is the engagement's deepest deliverable.
Take the audit first. Everything else is downstream.
Phase 0 is the cheapest, most reversible commitment in the engagement. Two to three weeks. A Brain Map, a Defect Inventory, a heuristic scorecard on your three highest-traffic surfaces, and a readout to the C-suite. No commitment to Phases 1–4 — though most agencies that finish Phase 0 commission them. The audit alone is a deliverable worth holding.
- Diagnostic
- Brain Audit · Defect Inventory · Heuristic Scorecards · Trust Canvas overlay. Two to three weeks. Fixed scope.
- Companion documents
- AUX Start Pack · 10 Heuristics · № 01 Agency Brains Design
- Start
- auxfirst.com/start — name the agency, scope the audit, hold the slot.
The agent is a user. The relationship is the product. Trust is the moat. The brain is where trust is actually earned.