How many minds
do you need?

1. Why one mind cannot hold everything

Ask the single marketing-agency Mind a simple question: what is your goal right now?

It cannot answer coherently, because it has been given goals that collide. “Maximize creative quality” pulls toward slow, crafted work; “ship daily to ride the trend” pulls the other way. “Match the client’s established brand voice” pulls toward consistency; “exploit what’s spiking in the market this week” pulls toward novelty. “Spend the media budget for reach” and “spend it for measured conversion” are in direct tension. A Mind is the unit within which priorities get ranked — and you cannot rank priorities that belong to different games.

This is not a context-window problem. You could give this Mind infinite context and infinite tools and it would be no better, because the failure is not “it ran out of room.” The failure is that it has no coherent answer to what it is trying to do. Attention has nothing to organize itself around. Learning cannot compound, because a win in creative tells it nothing about whether its media-buying judgment is improving. Experimentation is impossible, because when a campaign underperforms it cannot tell whether the copy was weak, the creative wrong, the targeting off, or the budget misallocated — the signal is smeared across too many domains to attribute.

Paper 2 defined a Mind as persistent cognitive state attached to an identity, with goals that bias attention. Live The hidden assumption in “one agent for everything” is that goals can be piled without limit into one attention stream. They cannot. A Mind has a finite focus, and that finitude is not a hardware limit to be engineered away. It is what makes the Mind able to prioritize at all.

Constraint is the feature.

2. The federation

Step back from software and the answer is two hundred years old. Why does a company not consist of one supremely capable person doing everything? Not only because no one has the hours — because no one has the coherence. The person who is optimizing this quarter’s cash flow is, in that moment, the wrong person to also be optimizing brand risk and database schemas, not for lack of talent but because those are different objectives that reward different attention. So organizations federate. A CFO with a coherent mandate, a CMO with another, a CTO with a third. They hold different goals, learn from different feedback, and disagree productively at the boundaries — and a coordinating function routes between them. The CEO does not hold all three mandates at once; the CEO routes.

This is the federation: bounded, autonomous focuses coordinating instead of one focus absorbing everything. The autonomy is load-bearing. Each node can master its domain, fail safely within it, and improve on its own clock precisely because it is not responsible for everything at once.

There is a tension worth naming honestly, because it looks like a counterargument. The theory of constraints tells us not to optimize the parts at the expense of the whole — optimize the system, not the silos. That is correct for a system that one mind can actually hold. But the world, across all time, is not a single system that a single optimizer can encompass. The only way to optimize something too large for one coherent focus is to compartmentalize it — to distribute the optimizing across many bounded focuses and coordinate them. Federation is not a retreat from whole-system thinking. It is the only form whole-system thinking can take once the system outgrows a single mind. The federation is the larger optimizer; its nodes are how it scales.

So the question “how many agents does the marketing agency need” is really: what are the coherent focus boundaries in this work, and how do they federate?

This also settles an argument the field had with itself. In 2025 two strong teams reached opposite public conclusions in the same week: Anthropic described building a multi-agent research system and argued that, in specific cases, many agents beat one — splitting work to keep each agent’s context clean; Cognition, the team behind Devin, published “Don’t Build Multi-Agents” and argued the reverse — that partitioning work partitions context, specialists make conflicting implicit decisions, and you should keep one coherent thread. Read closely, the two are made of the same material: one optimizes context by dividing it, the other by preserving it. Both are reasoning about a stateless agent, for which context is the only currency there is. A Mind changes the axis of the debate — the question stops being how to slice context and becomes where coherent focus begins and ends. The federation is the answer that falls out, and section 5 shows it dissolves the disagreement rather than refereeing it.

3. What a coherence boundary is

A Mind is warranted where there is a coherent domain of continuous judgment — a place where one goal can organize attention, where outcomes feed back into better decisions, and where mastery compounds over time. Three tests, together:

One goal can rank its priorities. Inside the boundary, when two options compete, the focus can say which matters more without appealing to some other domain’s objective. “Bolder concept or safer brand fit?” is answerable inside creative production. “Better creative or better media targeting?” is not — that question lives a level up, at coordination, because it spans two focuses.

Outcomes feed learning. The boundary is drawn so that when something works or fails, the focus can tell why within its own terms and get better next time. Creative production learns from how its work performs; that signal is clean because it is not entangled with media-buying or strategy choices.

Mastery compounds and evolves. The judgment inside the boundary is not a fixed procedure. It changes as the world changes — a new model drops, the market’s taste shifts, an ad platform that was best last quarter is overtaken. The focus continuously re-evaluates and reprioritizes. This is the property that most sharply separates a Mind from a tool. Paper 1 makes the same point: a Mind holds beliefs that change; memory and tooling hold stored capability.

Run these on the marketing agency and the boundaries appear:

• Creative production mastery — knows the copy, design, and video tools individually; knows which suits a brand spot versus a performance ad versus a social cutdown; continuously re-evaluates as new models arrive. One goal (the best creative execution of a given intent), clean feedback, compounding taste. A Mind.
• Media buying — the same shape over a different domain: a dozen ad platforms and targeting tools, evolving judgment about which channel and bid fits what objective. Orthogonal to creative; its own feedback loop. A Mind.
• Campaign strategy — the focus that decides what a campaign should be: this push wants a brand film, this one wants performance ads, this one a social series. It composes across disciplines without executing any of them. A coherent goal (the shape of the campaign), its own learning. A Mind.
• Market-trends intelligence — what is changing in the market and channels that affects what works. Its own clock, its own feedback. A Mind, and notably one whose learning is reusable across many clients.

And — the caution that keeps this from running away — the things that are not boundaries:

• “Open a new project in the design tool. File, New.” A fixed procedure. No goal to rank, no judgment that compounds, nothing that improves with experience. This is a skill the creative Mind uses, not a Mind.
• “Export at 1080p.” A capability in the tool layer. Not a focus.
• “Resize the banner to 300×250.” A task within creative production, not a domain of its own.

The test that separates the two: does this thing make judgments within a coherent domain and get better at them, or does it execute a defined step? Mastery is a Mind; procedure is a skill. If you find yourself drawing a Mind around “start a new file,” you have mistaken a microtask for a focus, and a federation of microtask-minds is the opposite failure — a swarm so finely divided it spends all its time coordinating and none of it mastering anything.

4. Skills, tools, and the line beneath the Mind

It is worth being exact about what lives inside a coherence boundary, because the federation only works if each Mind is genuinely holding a domain rather than a pile of procedures.

Inside the creative Mind there are many skills and tools: the copy models, the image and video generators, the editors, the layout tools. These are capabilities — Paper 1’s memory and tooling layer. They can be added, removed, version-controlled, swapped for better ones. They do not, individually, learn; they are used. What learns is the Mind around them: its evolving belief about which tool fits which job, reinforced every time a choice proves right and decayed when a better option appears. The tools live in the tool layer; the earned opinion about the tools is the Mind. Live

This is the crucial difference between a skill and a Mind, and it is exactly the difference your gut points at when you ask “is knowing when to use the tools just another skill?” No — because a skill is written once and refactored occasionally, while the knowledge of when and which is continuously adapted from outcomes. A fixed sequencing rule is a procedure. A living, reprioritizing judgment about sequence and selection is the work of a Mind. The moment “how to use these tools well” stops being a static rule and becomes a continuously-learned competence, it has become the focus of a Mind, not an entry on the tool layer.

So the layering, bottom to top: tools and skills (fixed, swappable) sit inside a Mind (a coherent focus that holds evolving judgment about them), and Minds are coordinated by a federation (routing across focuses). Stated in the vocabulary the rest of the system uses: a Mind holds its own cognition, uses the tools and sources it links to, and works with other minds. Holds is the focus; uses is the tool layer; works-with is the federation. Trouble comes from collapsing the wrong pair: treating a microtask as a Mind (over-splitting into a coordination swamp), or treating a whole mastery domain as a skill bolted onto one overloaded generalist Mind.

5. Coordination: the federation needs a router, not an emperor

If the marketing agency is a federation of focused Minds, something has to route among them. The strategy Mind decides a push wants a brand film; the creative Mind must then evaluate which of its tools can deliver that; if none can at the needed quality, that has to travel back so the strategy Mind can reconsider. Who holds that flow?

A coordinating focus does — but it is a router, not an emperor. Its job is not to do creative or media or strategy; it is to hold the overall intent ( “make this campaign land for this client”) and route each sub-goal to the focus that owns it, carry results between them, and reconcile conflicts at the boundaries. This is the CEO who does not hold the CFO’s, CMO’s, and CTO’s mandates simultaneously, but routes between them and owns the question of how they fit.

Two things keep this from quietly becoming one giant Mind again. First, the router does not absorb the specialists’ judgment — when it wants a creative decision, it asks the creative Mind, it does not decide for it. The moment the router starts overruling specialists from its own opinion, you have re-centralized into one incoherent focus. Second, the router’s own competence — which specialist to trust for which call, how to sequence them, how to resolve a standoff — is itself a coherent, continuously-learned focus. Routing well is a mastery domain in its own right, with its own feedback (did the campaign land?) and its own evolving judgment. So the coordinator is a Mind too, with a genuinely coherent goal: orchestration. It is not “the agent that also does everything.” It is the agent whose domain is the coordination. In progress · Phase 2

This is the resolution promised earlier to the Anthropic-versus-Cognition debate. Anthropic’s lead-spawns-subagents pattern is a federation with a router; Cognition’s warning — that partitioned agents make conflicting implicit decisions — is precisely what happens when the federation has no coherent router and the nodes are left to silently disagree. The fix Cognition reaches for (collapse back to one agent) over-corrects toward one overloaded focus; the fix the federation prescribes is a coherent coordinating focus that owns the boundaries. Spawn for focus; coordinate through a router; never let the router become an emperor.

6. The shared world keeps the federation honest

A federation raises an obvious worry: if these Minds are autonomous, do they drift into private, contradictory versions of reality — five specialists who no longer agree on the basic facts of the client they serve?

They do not, because of the architecture in Paper 2. Claims and the entities they concern live once, in a shared world layer. A Mind does not own a private copy of “this client’s audience skews young and expects fast, native content.” It owns an evaluation of that shared claim — what it means for its domain, how much it weighs, how it bears on its choices. Live The creative Mind reads that audience fact and concludes “punchy, native-feeling spots”; the media Mind reads the same fact and concludes “short-form social channels, mobile-first bids”; the strategy Mind reads it and concludes “lead with the social series.” One world, many coherent judgments about it. That is not drift. That is exactly the productive specialization a federation is for.

This is what makes federation safe where the naive “many agents” design is dangerous. The danger in spawning agents was always that each gets its own siloed memory and they fall out of sync. Here the facts are shared and singular; only judgment is distributed. So coordination is not a fragile re-narration of facts at every handoff (the stateless “telephone game”) — the facts are common ground. Coordination is the smaller task of reconciling judgments when they bear on a shared decision. The federation disagrees about what to do, never about what is true, and that is the healthy kind of disagreement.

It also reframes the real cost of getting the boundaries wrong. Over-split into too many Minds and you do not duplicate facts — the world stays single — but you fragment evaluation past the point of coherence, and you pay a cold-start tax on every Mind, since (Paper 1) a new Mind with access to a large memory store still knows nothing until it has engaged and built belief through use. Five Minds where two would do means three extra focuses each climbing its own learning curve from zero, coordinating more and mastering less.

7. How to engineer your own federation

This is meant to be buildable, so here is the procedure, not just the principle.

1. List the work as goals, not tasks. Write down what the system is trying to achieve, in objective terms — “produce creative that lands,” “buy media that converts,” “decide what each campaign should be,” “stay current with the market,” “coordinate the whole into a campaign.” Resist listing tools or steps yet.

2. Group goals that can be ranked against each other without leaving the group. Goals that trade off cleanly among themselves belong to one focus. Goals that can only be adjudicated by appealing to a different objective belong to different focuses. “Bolder or safer concept” stays inside creative. “Better creative or better media targeting” escapes to coordination — so creative and media are different Minds.

3. Confirm each group has its own feedback loop. A real focus learns from outcomes it can attribute to itself. If a candidate boundary can’t tell whether its own choices succeeded — because the signal is entangled with another domain’s — it is not yet a clean boundary; redraw it.

4. Demote the procedures. Everything inside a focus that is a fixed step — open the file, export the format, call the API — is a skill or tool in the tool layer, not a sub-Mind. If a “specialist” you sketched turns out to do one defined procedure and never develops judgment, it is a skill. Fold it in.

5. Name the router. Identify the coordinating focus whose domain is orchestration itself. Give it the overall intent and the authority to route — not to overrule. Make sure its judgment (whom to trust, how to sequence, how to break ties) is allowed to learn.

6. Check the count from both ends. Too few Minds: do any of your focuses contain colliding goals that cannot be jointly ranked? Split them. Too many: are any of your Minds drawn around a single procedure, or do they coordinate more than they master? Merge them. The right number is the number of genuinely coherent, independently-learning focuses the work contains — no fewer, because fewer means incoherence; no more, because more means a coordination swamp.

For the marketing agency this yields, roughly: a campaign-strategy Mind that decides what each campaign should be; a creative Mind and a media-buying Mind that master their respective tool domains and execute intent; a market-trends Mind tracking the ecosystem (and reusable across many clients); and a coordination Mind holding the campaign intent and routing among them. Per-client confidentiality, on inspection, is not its own Mind (“serve each client without leaking across competing accounts” is handled by scope, not by a separate focus), governed by the audience and disclosure machinery of Paper 3 rather than split into its own swarm. Planned · Phase 2 The exact list is arguable — and it should be argued, on these tests, which is the point. The framework is not the answer; it is how you find your answer.

8. What this rests on, and what is still being built

This series describes a product architecture, not a fictional future, so the boundary matters here.

Live and load-bearing for this argument: per-identity Minds for humans and agents, the separation of claim from per-Mind evaluation, the shared world layer that lets many Minds hold one set of facts, source evidence, the attention and consolidation loop that gives a focus its goal-biased prioritization, skill nodes (the Mind’s evaluated belief about a capability, distinct from the capability itself, which lives in the memory and tooling layer), and multi-owner agents. Live These are what make “a Mind is a coherence boundary” a thing you can actually instantiate rather than a metaphor.

Built on that substrate but not yet turnkey: the federation as a first-class pattern — a coordinating Mind routing among specialist Minds, with the boundary-reconciliation that implies — is the architectural commitment this paper makes, not a finished button. In progress · Phase 2 The layered scoping it leans on for cases like per-client confidentiality (Paper 3’s shared core plus per-user perspective) is phase 2. Planned · Phase 2 And the earned judgment each specialist holds about its tools depends on the conversion loop running well over many uses; the substrate is live, well-calibrated mastery is what it matures into. Read sections 3 through 7 as architecture on a real foundation, with the honest note that composing the federation cleanly is the work in front of us.

9. What it trades

No architecture is free, and the federation has real costs pointing back the other way.

Coordination is not free, and a bad router is worse than none. Federation moves the hard problem to the boundaries — sequencing, conflict resolution, knowing which specialist to trust. A weak coordinating focus produces exactly the conflicting-implicit-decisions failure Cognition warned about. The federation is only as good as its router.

More minds can cost more compute. A federation of focused Minds may consume more total compute than one giant Mind would, but the giant Mind does not actually work for incoherent goals. The honest version: federation trades raw efficiency for coherence, learnability, and the ability to fail safely in one domain without poisoning the others. Usually worth it; not always; measure it.

Boundaries can be drawn wrong, in both directions. Too few and you reintroduce incoherence inside an overloaded Mind. Too many and you drown in coordination. The tests in section 3 reduce this risk; they do not eliminate the judgment.

The finite Mind is a commitment, not a limitation to route around. The whole argument rests on keeping each Mind’s focus bounded. The pressure, always, will be to let a successful specialist accrete adjacent responsibilities until it is quietly overloaded again. Holding the boundary is ongoing discipline, not a one-time design choice.

These are the costs of treating “how many minds” as a question about coherence rather than capability. They are smaller than the cost of building one Mind that cannot tell you what it is trying to do.

10. Read together

The first three papers built the foundation and can be read in any order. This one sits on top of all of them.

Your AI has amnesia established the Mind — persistent cognitive state per identity, with goals that bias a single coherent stream of attention. Two minds, one room established the layers within a Mind — shared core, per-user perspective, per-room context — that scope relationships and disclosure inside one focus. From memory to cognition established the maturity ladder beneath every Mind — that memory is not cognition, and that a fixed procedure is not a focus. This paper spends all three to answer the design question they imply: not what a Mind is, but how many you need, and where the lines fall.

The series has run the same move four times. Memory is not cognition. One mind in the room is not two. Stored material is not understanding. And now: one mind holding everything is not intelligence. Each time, the industry has reached for the bigger single thing — more memory, more context, more tools, one ever-larger agent — and each time the answer turned out to be structural rather than bigger.

So: how many minds do you need to build the marketing agency? As many as the work has coherent, continuously-learning focuses — and they are arranged as a federation, each bounded enough to know what it is trying to do, coordinated by a router that owns the whole intent without absorbing the parts. Not one mind straining to be everything. Not a thousand minds split down to microtasks. A federation of focused minds, which is the only architecture that has ever scaled a complex system — because the constraint that each focus is finite is not the flaw in the design. It is the design.