The Agent Organization

Challenge

One assistant working in one repository is a productivity tool. Nineteen codebases is an organisation, and organisations fail for organisational reasons: work happens out of order, nobody checks the work, and confident claims go unchallenged because the thing making them sounds sure of itself.

Language models are very good at sounding sure. That is the failure mode to design around.

What I did

• Split the work into roles and ran it in stages, so no agent is asked to plan, execute and judge the same task.
• Added independent verifier agents whose only job is to check work they did not do. A verifier that also writes code will approve its own.
• Required evidence tags on every claim: VERIFIED for something checked directly, REPORTED for something an agent was told, INFERRED for something concluded. A claim with no tag does not count.
• Wrote standing rules that hold across every codebase, so behaviour does not have to be re-specified per project.
• Managed it the way you would manage a team: clear scope per role, a hand-off protocol, and a review step that can send work back.

Result

• A 20-agent workflow operating across 19 codebases.
• Verification carried out by agents with no stake in the result.
• Every claim carrying a tag that says how strongly it is known.
• A method that transfers. The same structure runs the research work in Deep Research OS.