The Ground Conditions Changed — Reexamine the Organization
A multi-agent team is not just division of labor, handoffs, and context isolation. The higher-order question is whether it helps the system get closer to the truth.
Treat an agent as a worker, and the engineering objects — role, capability, track record, environment, responsibility — snap back into their proper places. Assign agents job grades, departments, meetings, and reporting lines copied from a corporate org chart, and you are merely replicating the corporate shell.
Human resources is the backbone metaphor of this book, but a metaphor is not a substitute for design. At the team level, the most dangerous shortcut is company simulation: importing the shell of a human organization and mistaking it for governance.
The Cost Structure of Organization Has Changed
Human organizations didn't grow into their current shape out of nothing. The hierarchy, the reporting, the approvals — much of it comes from the limits of the human brain: how much information one person can process at once, how many roles they can switch between reliably, how much context they can hold, how much psychological pressure they can bear. Much of what passes for management wisdom is patchwork over these limits.
Once agents enter, those limits no longer hold in the same form.
First, the window gets bigger and the middle layers weaken. Human organizations need thick middle layers to compress information: the front line touches the facts but can't digest the whole, the top makes tradeoffs but sits too far from the facts, and the middle compresses facts into reports and details into conclusions. An agent that can reliably process a very large body of material can read the traces, the raw material, the deliverables directly — no waiting for layer after layer of reporting to flatten the world. The organization gets flatter: facts flow faster to where judgment happens, and judgment flows faster back to where the work happens.
There's a premise here that "bigger windows" tends to obscure: being able to digest a huge pile of material is itself a tier of capability. Stuff a weak model's context full and it won't digest it — it will get dragged off course by noise, and you'd have been better off letting a middle layer pre-digest. The middle layer's job of compressing the world into conclusions hasn't disappeared; you've just pretended it has. Flattening has a precondition: the role must be staffed with a model of sufficient tier. Cut the middle layer without that, and you haven't eliminated it — you've dumped its workload on someone who can't carry it.
Second, roles can be generated on demand. Humans struggle to keep roles cleanly separated — review your own draft right after writing it, and the review gets pulled by the preferences you wrote with; assess the risks of a plan you just proposed, and the assessment reflexively finds reasons for the plan. Agents carry no such baggage. A lead agent can spin up a critic agent, give it only the necessary material, and withhold the lead's self-justifications; it can spin up a fact-checking agent that starts from the raw material, not from the conclusion. These sub-agents don't need to persist, don't need a full track record, don't need a headcount. Each is simply one clean environment for judgment.
Third, transfers are cheap, so generalists are more common. The model itself has broad general knowledge, skills can be loaded on demand, tools can be granted per role, and a track record can be distilled and migrated. A human employee changing jobs must relearn, readjust, rebuild their environment; when an agent changes roles, much can be copied, trimmed, and recombined. It won't necessarily be competent right away, but the transfer costs far less.
Fourth, work records are more complete. Human organizations lean on reports to reconstruct work because the real process has already scattered: meetings, verbal commitments, hallway conversations, on-the-spot judgment calls are nearly impossible to reassemble. Agent systems are different. Tool calls, intermediate artifacts, contexts, failed paths, and rework can all be kept. If the trace can reconstruct the process, don't invent a ritual where the agent narrates how hard it worked.
Don't Wrap Agents in a Corporate Shell
When the ground conditions change, you don't copy the company over as a template.
Company simulation usually looks like this: giving an agent team managers, supervisors, meetings, reports, approvals, daily standups. It looks complete, but it copies only the form. Human organizations need those rituals because people forget, conceal, fight over credit, and dodge responsibility — and because information moves slowly and gets distorted along the way. A process the trace has already recorded doesn't need a weekly report performed upward. An action the permission system already forbids doesn't need layers of verbal sign-off. A temporary role destroyed on completion doesn't need a career ladder.
Worse, the shell drags the failure patterns of human organizations in with it: hierarchy substituting for facts, reports substituting for work records, a feeling of process substituting for responsibility.
A truth-seeking team cannot start from organizational shape. It must start from the evidence path: where do the facts come from, which judgments do they pass through, where could they be contaminated, who can veto, who can bear the consequences. A corporate shell answers none of these questions. It only makes the system look orderly.
Audit Every Role
Treating agents as workers does not mean more roles is better.
A role is an engineering object, not organizational decoration. It needs explicit responsibilities, permissions, deliverables, and escalation conditions. But slicing too fine is itself complexity. Take work a generalist agent could finish in one working environment, split it into three roles, and you get three deliverables, three rounds of context compression, three responsibility boundaries — the split brought no sharper judgment, only more interfaces.
Many roles were carved out historically not because of task structure but because humans couldn't cope: someone does nothing but summarize because others can't read all the material; someone does nothing but coordinate because information moves slowly; someone does nothing but approve because the organization doesn't trust the front line. Copy those roles to agents as-is and you import the complexity that human limits created.
So every split deserves the same interrogation: does this division come from task structure, or from human bandwidth? Does it protect the independence of fact-finding, or just manufacture a sense of organization? Does it reduce risk, or manufacture the appearance of sign-off? Is the handoff cost smaller than the gain from splitting? Does it make facts clearer, or responsibility blurrier?
Delete the fake roles; keep the real boundaries. Real boundaries usually appear in three places: where the task structure differs, where the risk level differs, and where a judgment of fact needs independence.
Design from Constraints, Not from the Shape of a Company
A multi-agent team should be designed from its constraints. If the problem is division of labor, design role boundaries. If the problem is handoffs, design deliverables and acceptance criteria. If the problem is risk, design delegation and escalation paths. If the problem is quality, design evals and work records. If the problem is responsibility, design a final point of accountability.
Human organizations have these constraints too, but the constraints stand on collaboration itself, not on the corporate shell. The company is merely one historical form that happened to carry them. An agent team can borrow the experience without copying the shell.