Team-Forming Constraints
Only two things justify splitting work across multiple agents: the task exceeds what one agent can hold at quality (capacity), or parts of it must be done by different minds because they poison each other (separation of concerns). A multi-agent design that answers neither is not a team — it is just more agents.
That negative claim is the sharp end. The vault already holds that agent count is a design variable (Bounded Fan-Out) and that more agents ≠ better (Agentic Simplicity). This page states the positive criterion those pages leave implicit: there is a short, closed list of reasons to form a team at all, and a proposed roster should have to name which one it answers.
Limit 1 — capacity. A single agent cannot absorb unlimited spend, because everything it reads and does piles into a context window; as the window fills, quality drops (the source notes this holds "even with new techniques like autocompaction"), and the model must eventually delegate or abandon the task. A team is how you spend more tokens than one agent can usefully hold — which is the mechanism the source attributes to Anthropic for why an agent team beat a lone frontier model, and which connects the capacity limit directly to Repeated-Sampling Scaling's token-spend finding. The source flags this limit as the moving one: what a single agent can hold keeps growing, so "it's hard to know what to assign to an agent when agents keep getting smarter."
Limit 2 — separation of concerns. Some work has parts that must be done by different minds not because one mind lacks the skill, but because the parts poison each other. This limit is stable — it predates AI by centuries and has nothing to do with capability:
- "The auditor who also kept the books isn't a worse auditor. He's just not an auditor at all."
- Peer review only works because the reviewer didn't write the paper.
- Your bank will not let the person who enters a payment be the person who approves it — "not because they're dishonest," but because reliable work has always been built from checks and balances made of multiple minds.
What agents add to this old trick is genuinely new: you cannot unknow something, but you can start a mind that has never seen it. You have read your own product page a thousand times and will never see it as a stranger does; you can now instantiate a reader who hasn't. The source's framing — "fresh eyes on demand for the first time in history" — is the strongest available argument that Context-Independent Review is durable structure rather than a quirk of current models. Where it pays best: genuine conflicts of interest you want balanced, and anything that needs reviewing twice to be trusted (a contract, a draft, a plan).
The two limits map onto two of Agent-Shape Triage's four questions (size and separation of concerns), which is not a coincidence — the triage test is these constraints made into a pre-flight estimate.
Claims
- Every team-of-agents design that works answers one of exactly two problems — capacity or separation of concerns; everything else is just more agents. principle — durable, and stated as a closed list. Its value is as a falsifiable gate on roster proposals: name the limit you're answering, or don't form the team. The source offers it as a design rule rather than a survey finding, so the closedness is an assertion — but a productive one, and the vault has no counterexample.
- Some work must be split across minds because the parts poison each other, not because one mind lacks the skill. principle — durable and pre-AI: audit independence, peer review, and segregation of duties are centuries-old instances. This is the limit that does not dissolve as models improve, which makes it the more reliable of the two.
- You cannot unknow something, but you can start a mind that has never seen it — fresh eyes on demand. principle — durable: contamination is a property of having-been-exposed, so the only remedy is a reader who wasn't, and that is now instantiable on demand. This is what agents add to the old checks-and-balances trick, and it is the general form of Context-Independent Review's code-scoped argument.
- The capacity limit is a moving target; the separation-of-concerns limit is not. principle — durable, and the practical consequence is that a team justified purely on capacity has an expiry date while one justified on separation of concerns does not. The source names the moving target as what makes the problem space hard: "it's hard to know what to assign to an agent when agents keep getting smarter."
- A single agent cannot absorb unlimited spend: context fills, quality drops even with autocompaction, and the model must delegate or abandon. observation — the mechanical basis for limit 1, and the reason capacity is a limit rather than a preference. See Context Compression, Context Smart Zone.
- The video states a team of agents is how you spend more tokens than one agent can usefully hold, and attributes this to Anthropic as the explanation for its multi-agent research system beating the lone frontier model. observation — the source's attributed claim; check-worthy. It is what links this page's capacity limit to Repeated-Sampling Scaling's token-spend finding.
- Reach for a team where there are genuine conflicts of interest to balance, or where something needs reviewing twice to be trusted — a contract, a draft, a plan. best practice — context: work whose cost of being wrong exceeds the cost of a second independent pass. The source's own scoping; it is a contingent judgment about stakes, not a rule, and low-stakes work shouldn't buy the second mind.
- You might be running a multi-agent system without choosing to: a large enough task handed to a current chat product may cause it to delegate internally. observation — the source names ChatGPT 5.6; check-worthy. The point he draws is about intent, not mechanism — if the delegation is going to happen, decide it deliberately.
Related
- Agent-Shape Triage — the pre-flight test these two limits become: size is the capacity limit and separation of concerns is the second, turned into questions an operator can answer in a minute.
- Bounded Fan-Out — the next question down, and a clean division of labour: this page says whether to form a team and why; that page says how many agents once you have. "Name the limit you're answering" is the qualitative gate; cardinality is the quantitative one.
- Agentic Simplicity — "more agents, more autonomy, more compute do not always mean better outcomes" is the disposition; this page supplies the positive test that makes it actionable.
- Context-Independent Review — the second limit's best-developed instance in the vault, scoped to code review and argued from authorship anchoring. This page supplies the general principle (you cannot unknow) and the pre-AI lineage (audit, peer review, segregation of duties) that page lacks.
- Authority-Independent Verification — separation of concerns pushed up the org chart: not only must the checker be a different mind, no rank is exempt from being checked.
- Cross-Model Independence — a third reason two minds beat one that this page's list does not cover: correlated blind spots. Worth noting against the "exactly two" claim — lineage independence buys something neither capacity nor a fresh context does, so the closed list may be closed only over work-shape reasons, not over all reasons to split.
- Repeated-Sampling Scaling — the capacity limit's economic twin: a team is how you spend past one window, and token spend is what the source says predicted good runs.
- Agent Communication Topology — once the limit justifies a team, whether the parts talk is the next design choice; note that limit 2 often requires they don't.
- Step Isolation — the same hide-what-would-bias-you move inside a single skill; separation of concerns is that principle at roster scale.
- Role-Typed Agent Roster — the staffing form: roles exist because concerns must be separated, so the roster is limit 2 made concrete.
- Context Compression — the mitigation that pushes the capacity limit outward rather than answering it with a team; the source's "even with autocompaction" is the caveat that it doesn't push far enough.
- Agent Task Graph — "more agents ≠ better; what matters is coordination toward a common goal" is the coordination-side statement of the same discipline.
- Distillate: 1.6M agents registered for OpenClaw and did NOTHING.