Adding more agents is easy.

Reviewing more agent output is hard.

The coordinator exists because those two facts collide. It is the part of the system that turns many candidate futures into one reviewable decision without letting review bandwidth collapse.

Attempts Are Supply

An agent can produce a patch, test, diagnosis, benchmark, question, fixture, or partial proof.

None of those is automatically the outcome.

worker A: patch plus weak test
worker B: diagnosis plus no patch
worker C: broad rewrite
worker D: focused fixture
worker E: stale but useful evidence

If every output arrives as a transcript, the coordinator has to reconstruct the whole run by hand. More agents can then slow the system down.

The useful unit is the reviewable bundle:

candidate
changed regions
patch summary
evidence
risk
stale head
missing proof
recommended route

That bundle is small enough to compare.

Review Bandwidth Is Demand

Agent output is supply.

Review bandwidth is demand.

When supply grows faster than review capacity, quality falls even if worker count rises. The answer is not fewer agents by default. The answer is shaped output, better compression, risk routing, and a willingness to stop work that no longer changes the decision.

Context Is A Resource

Coordinator work is also context budgeting.

Workers should not all receive the whole world. They should receive the smallest slice that lets them produce useful evidence.

worker A: parser fixture and expected output
worker B: public API surface and type gate
worker C: failing browser step and DOM probe

That makes worker output easier to compare. It also keeps failures local. If a worker had the wrong context, the coordinator can fix the slice instead of treating the result as mysterious.

Route By Review Cost

Not every candidate deserves the same review path.

low risk + strong evidence -> admit
low risk + missing evidence -> rerun
high risk + strong evidence -> human review
high risk + weak evidence -> split or block

That routing protects human attention. The human should see the cases where judgment matters, not every case where an agent was verbose.

Waves

Large agent systems move in waves.

One wave explores. Another implements. Another verifies. Another reviews what the previous waves missed.

A useful coordinator does not keep every worker running forever. It collects when enough evidence exists to make the next decision.

Long-tail work is not free. The coordinator should ask whether a slow worker can still change the decision. If the lane is duplicate, stale, or unlikely to produce new evidence, stopping and collecting can be the correct move.

The Backlog Changes

A coordinator backlog is alive.

When an agent finds a missing oracle, that becomes a task. When a patch is stale, that becomes a rebase task. When two workers disagree, that becomes a comparison task. When the system needs a human answer, that becomes a typed blocker.

new task: build focused fixture
new task: rerun worker patch against current head
new task: compare two diagnoses
blocked: user must choose public API behavior

This is why coordinator work belongs in the run graph. The backlog is not just a to-do list. It is part of the history of why the final state exists.

The Mental Model

Agents create candidate futures.

The coordinator spends limited review bandwidth on deciding which future is supported by evidence.

Good coordinator work makes the system faster without making it less reviewable. It lets many workers move at once while keeping the final decision small enough for a human to understand.