Trust, but Verify. Then Verify the Verify.

On a long task, one agent is not enough of a check.

2026-10-03

Idea

On a long task, one agent is not enough of a check. We have been spawning three, numbered from zero: agent 0, agent 1, and agent 2.

Problem

A long task is a chain. Steps get skipped, softened, or marked done.

• If agent 0 grades its own work, it agrees with itself.
• Two agents can share the same shortcut: one does less, the other calls that enough.
• The chain still looks finished.

Solution

Three agents. Three jobs. No shared grade.

1. Agent 0 does the job.
2. Agent 1 checks that the job was actually done.
3. Agent 2 does not redo the job, and it does not grade agent 0 on its own. It checks that agent 0's result and agent 1's verification agree.

The third pass is spent on whether the first two agree, not on a third trip through the work.

Theory

A check that shares context with the worker inherits the worker's misses. A second pass from the same agent is not a second opinion. Two separate checks can still land on the same shortcut. Agent 2 is there to catch that, by looking at the two results apart from both of them.

A fourth agent, agent 3, can check that agreement again. Useful, and it costs another run. So far three is the cut that stays worth the cost. Past that, the spend is on checking the check of the check.

Practice

Spawn 0, 1, and 2 at the start, not after the chain has already been called done.

• Agent 0 owns the artifact.
• Agent 1 owns a pass or a fail against that artifact, with the miss named.
• Agent 2 owns a pass or a fail on whether those two agree.
• A fail sends the chain back.
• A pass names what was checked. It is not a note that the work seemed fine.

These prompts are a basic demonstration of the idea, and a practice that has been working. They are not a script to keep pasting unchanged. As agents get stronger, the wording should change. The split should not: one does the job, one checks the job, one checks that those two agree.

Fill-in prompts: Spawn prompts.