Agency Brains Design.01

The mistake is shared across every holdco AI initiative we have seen and every roadmap deck circulating at every level of the agency stack. It is the assumption that once you give an LLM access to the right knowledge, it will reason correctly. That is not how reasoning works in either machines or organisations.

Every meaningful question an agency answers has a required reasoning type. Compliance questions need rule-based, deterministic reasoning — outputs that are reproducible and citable to a specific clause. Effectiveness questions need causal reasoning — outputs that distinguish lift from coincidence and that disclose their counterfactual. Forecasting questions need probabilistic and statistical reasoning — outputs with confidence intervals, not single numbers. Creative questions need stochastic and generative reasoning — variation is the point, not the defect.

A generative LLM, reasoning over a knowledge corpus, applies generative reasoning to all of them. It produces a fluent paragraph for a forecasting question. It produces a fluent paragraph for a compliance question. It produces a fluent paragraph for a causal question. The fluency masks the fact that no two of those paragraphs were generated by the same kind of cognitive process the question actually required.

This is not a model-quality problem. Better models do not change the reasoning type. A GPT-5 stochastic-generative answer to a compliance question is still a stochastic-generative answer to a compliance question. The defect is one layer below the model. It is the absence of a declared reasoning type, an absence of routing, an absence of a typed cognitive architecture for the agency to stand on.

What the defect produces, today

We see it in every audit. Account teams using ChatGPT-class tools to "stress test" claims — receiving fluent passes for claims that are not in fact substantiable, because the tool is not running a rule engine, it is running text completion. Trading desks using LLMs to write performance narratives — receiving causal-sounding sentences for outcomes the data does not in fact support. Strategists using LLMs to size audiences — receiving plausible numbers with no confidence band and no traceability. Compliance reviewers using LLMs as triage — receiving recommendations with no jurisdiction tag and no rule reference.

Each of these is a misapplied brain. None of them is solved by adding more agents on top. Agents amplify the underlying cognitive architecture; they do not correct it.

What the unlock looks like

The agency declares a brain library. Each brain has a typed reasoning method — chosen from the standard taxonomy below — and a manifest that specifies its knowledge sources, output contract, and trust signature. A meta-brain router sits in front of every AI surface in the agency, decides for each request which brain is qualified to answer, and routes the call. Every output carries the brain ID it came from and the reasoning type it used.

Three things change. Audit becomes possible — every output is traceable to a declared brain with a declared method. Trust becomes localised — a failing brain can be fixed without taking down the entire AI surface. Capability becomes additive — new question types are answered by adding a new brain, not by retraining the world.

Every brain in the library is one of these reasoning types or a declared composite of two. The full taxonomy below — the same one a systems theorist would write down — exists so that no agency project can hide its reasoning type by accident. The declared type is the contract.

The six core types every agency brain library starts with

Of the fourteen, six pull most of the weight in a media or advertising agency. Deterministic, rule-based, probabilistic, causal, heuristic, and generative. A starter brain library should cover all six, then extend into statistical, Bayesian, and agent-based as the business asks for it. Fuzzy and stochastic typically appear as composites with another type, not as standalone brains. Chaotic and emergent appear most often as labelling primitives — the brain that recognises when no reliable prediction is possible.

The meta-brain is the single most important component, and the one most likely to be skipped. Without it, every request still defaults to whatever AI surface received it — which means default-stochastic, default-generative, default-untyped. The meta-brain is the place where the agency decides, per request, what kind of reasoning is appropriate.

Five signals the router classifies on

The router does not need to understand the question deeply. It needs to read five signals well enough to pick the right brain. These five are sufficient for the vast majority of agency traffic.

1. Reproducibility requirement. Must the same input produce the same output? If yes → deterministic or rule-based. If no → the answer space is open.
2. Audit and disclosure requirement. Will this output be cited to a client, regulator, finance team, or court? If yes → the brain must carry a typed trust signature with rule references or methodology disclosure.
3. Causal claim. Is the response claiming that X caused Y? If yes → causal brain only; correlation-based brains forbidden.
4. Forecasting horizon. Is the answer about a future state? If yes → probabilistic or statistical; never generative without confidence intervals.
5. Variation requirement. Is the goal one canonical answer or multiple options? If the goal is variation → generative/stochastic and variants are always plural.

How the router itself reasons

The meta-brain is itself rule-based with a small heuristic layer. It is not an LLM "deciding" what brain to call — that would just re-introduce stochastic routing into a system designed to eliminate it. The router is an explicit decision tree, version-controlled, auditable, with logging on every routing call. When the rules cannot resolve, it falls back to a declared default brain — typically Brief Brain — and flags the request for human routing review. The router is the only place where rule-based reasoning is non-negotiable.

What the router returns alongside the answer

Every response from the system carries four pieces of metadata: the brain ID that answered, the reasoning type used, the knowledge sources consulted, and the trust signature (citation list, confidence interval, methodology card, or rule references — whichever applies for that type). This is the audit substrate. The agency can finally answer the questions a holdco general counsel, a CFO, or a client procurement team will eventually ask: who said this, on what basis, with what data, under what method.

The manifest is the contract a brain signs with the agency. It is version-controlled, reviewable, and required before a brain can be invoked by the router. If a system in the agency reasons and does not have a manifest, it is shadow cognition — exactly the state the offering is designed to end.

The manifest format is minimal on purpose. The point is the discipline of declaring, not the bureaucracy of the form.

# brand-brain.manifest.yaml
brain_id:          brand-brain
version:           2.1.0
reasoning_type:    [deterministic, rule-based]

purpose:           "Enforces brand voice, claim policy, asset usage rules."

knowledge_sources:
  - mcp://brandbook.agency.local/v4
  - mcp://claims-policy.agency.local/2026-q2
  - mcp://asset-rights.agency.local/current

reasoning_engine:
  type:            rule-engine
  implementation:  drools / OPA / custom
  llm_role:        "text classification only, never adjudication"

output_contract:
  shape:           pass_fail_with_citations
  required_fields:
    - verdict             # pass | fail | needs_review
    - citations           # list of brand-book § references
    - jurisdiction        # market / region / brand
    - rule_version        # exact rule set tag

trust_signature:
  type:            rule_reference
  verifiable_by:   "any human reading the brand book"
  audit_log:       required

router_tags:
  invoke_when:
    - "output cites or implies brand position"
    - "reproducibility required"
    - "claim substantiation requested"

owner:             Brand Council
review_cadence:    quarterly
last_reviewed:     2026-04-18

The manifest is the document that makes a brain real. A brain without a manifest is not part of the system. The router does not call brains that are not registered. Outputs that come from elsewhere are flagged as unmanifested cognition and surfaced to the brain council for resolution.

The five things every manifest must declare

1. Reasoning type — chosen from the taxonomy. Composites of two allowed; single-type preferred.
2. Knowledge sources — typed MCP server URIs, dataset paths, model artefacts. No "the model knows" handwaves.
3. Output contract — required fields, response shape, what every response must include for the type to be honoured (e.g. confidence intervals for probabilistic, counterfactual for causal, citations for rule-based).
4. Trust signature — how a downstream human or auditor verifies the output is what the brain claims. Different per type. Rule-based: citations to rules. Causal: methodology card. Probabilistic: confidence interval. Generative: variant count and diversity disclosure.
5. Router tags — declared invocation conditions, used by the meta-brain. The brain says when it should be called; the router enforces.

The stack below is the starter library. Six are core — they cover the audit-critical, fiduciary-critical, and effectiveness-claim surfaces. Four extend into modelling, simulation, and optimisation. An agency that builds these ten well has done more for its AI capability than another that has deployed forty agents on top of an undifferentiated LLM.

How the ten compose

The brains compose in three useful patterns. The Brand Brain and Compliance Brain run as gates — every output that touches client-facing surface area passes through them before it leaves the system. The Effectiveness Brain runs as the single authority for any causal claim — case studies, capability decks, results paragraphs in pitches. The Creative Brain feeds the gates — it generates plurals, the gates reject the non-compliant variants, the survivors enter the work.

The pattern matters because the alternative — every brain answers in isolation, the agency hopes it all lines up — is what most agencies have today. Composition is the second discipline; declared types is the first.

The brain layer is not a new product the agency buys. It is a declared cognitive contract the agency authors and routes through. It sits between the AI surfaces — chat copilots, internal agents, briefing tools, agentic buying systems — and the MCP knowledge servers the agency has already built or will build.

What every AI surface in the agency must do, after

Every AI surface — every copilot, every agent, every internal tool — calls the meta-brain router on every request. The surfaces become routing clients, not reasoning endpoints. They format the request, send it to the router, receive the typed response, and present it. They stop being the place where reasoning happens. Reasoning happens in the brain layer.

This is a small code change and a large governance change. The code change is one HTTP call. The governance change is that no AI surface is allowed to "just call an LLM directly" for any production-facing question. That habit ends the day the brain layer goes live.

What MCP servers do, after

MCP servers continue to do exactly what the previous piece in this series described — they expose typed read access to corpora. The brain layer is the consumer of MCP, not a replacement for it. Each brain in the library declares which MCP servers it reads from, in its manifest. The MCP servers themselves remain agnostic — they serve any client that holds the right scope. The plumbing stays. The cognition gets named.

What agentic systems do, after

Agentic systems — Account Agents, Buying Agents, Brief Agents, others — call the brain layer for every reasoning step. They become orchestrators of brain calls plus actions. When an agent needs to "decide if this audience is large enough" — that goes to Audience Brain. When the same agent needs to "draft three alternative cohort framings" — that goes to Creative Brain. When the agent needs to "check whether this targeting is compliant in DE and FR" — that goes to Compliance Brain. The agent stops being a single fluent reasoner and becomes a typed-call orchestrator.

The audit substrate this produces

Every response in the agency carries: which brain answered, what reasoning type was used, what knowledge sources were read, what trust signature was returned. Aggregated over a week of operation, this gives the agency something it has never had before — a verifiable map of how its AI is actually reasoning across the business. That map is what the CTO presents to the holdco risk committee. That map is what the CFO uses to argue spend. That map is what the agency uses to defend itself when something goes wrong.

The engagement is structured into five phases. Phase 0 is a diagnostic — small, fixed-scope, accessible. Phases 1 through 3 are the build. Phase 4 is the operating model that lives inside the agency afterwards. The intent is to leave the agency self-sufficient at the brain layer, with a council that owns the manifests and a typed cognition stack it can extend without us.

Agency Brains Design is priced as a constitutional engagement — it touches everything underneath, the work is high-leverage and high-judgement, and the deliverables are durable enough to outlast any individual campaign or AI tool. The shape below is the indicative structure; we tailor for each holdco.

Why the price holds

Three reasons. First — the deliverable is durable. The brain spec, the manifests, the router, the council are good for years. They are not bought again next quarter. Second — the alternative is the cost of the defect. A causal claim made by a stochastic-generative LLM and put in a client deck is a fiduciary exposure that does not exist for an agency running an Effectiveness Brain with a published methodology card. Third — the engagement is the unlock. Every subsequent AI initiative the agency runs is faster, safer, and cheaper because the constitutional layer is in place.

The build order we defend

Brand Brain first. The reasoning is that brand consistency gates almost every external-facing output, and a deterministic-rule-based brain is the easiest to declare a manifest for and the easiest to demonstrate value with. Once Brand Brain is routing, the agency has a working router, a working manifest format, and an audit trail visible to leadership — the three prerequisites for everything else.

Brief Brain second. Briefs are where strategist time is most leverageable and where heuristic-fuzzy reasoning is most obviously the right type. The team sees the difference daily, the case for the rest of the library writes itself.

Effectiveness Brain third. This is where the trust dividend is largest and the political moment is sharpest — if causal claims in capability decks are coming from a typed brain with a methodology card, the agency's effectiveness story acquires a layer of defensibility that is rare in the industry. This is where pitches start being won on substrate, not just slides.

What happens to agencies that don't do this

Three things, in our prediction. First — the audit gap closes against them, not for them. By the end of 2026 we expect at least one prominent agency to be embarrassed in public by a causal claim that turned out to be LLM-generated correlation, and the holdco that follows will face a contractual indemnity demand it does not have the substrate to answer. Second — the trust gap with clients widens. Procurement teams at major advertisers are already asking what reasoning method is behind a recommendation; agencies that cannot answer in typed terms lose ground to those who can. Third — every internal AI initiative compounds the defect. More agents on top of an untyped brain is more surface area for the same underlying problem.

The window for the constitutional engagement is now. The agencies that build this layer in 2026 will be quietly running a different business in 2028 — one whose AI outputs are defensible, whose agents are localisable, whose stack is composable. The agencies that do not will be replatforming under pressure.