Research and Policy Analyst · EI MAR Automation

Who does what

One AI agent builds, a second is asked to find faults in the work, and a third, on a different AI system, audits the results independently. I plan, rule and approve.

Explanatory diagram of the rolesBuild · challenge · audit · decide
Who does what: three AI agents and the person who decidesI plan, rule and approve: I agree each plan and give my approval, run the programs on the real data, rule on every difference and approve the result. The builder agent writes each program and its tests from the agreed plan and fixes the faults that are found. The reviewer agent, a separate agent, reads the work fresh and is asked to find faults, ranked blocker, major or minor. The second builder, a different AI system working separately, audits the results cell by cell. The plan and my approval pass to the builder; faults and differences found come back to me. ME I plan, rule and approve I agree each plan and give my approval. I run the programs on the real data, rule on every difference and approve the result. the plan, and my approval faults found differences found BUILDER AGENT An AI coding agent Writes each program and its tests from the agreed plan, and fixes the faults that are found. REVIEWER AGENT A separate agent Reads the work fresh and is asked to find faults, ranked blocker, major or minor. SECOND BUILDER Another AI system Working separately, it audits the results independently, cell by cell. the work goes over, faults come back the results are audited independently, cell by cell

Me

I plan, rule and approve

I agree each plan and give my approval. I run the programs on the real data, rule on every difference and approve the result.

↓

Builder agent

An AI coding agent

Writes each program and its tests from the agreed plan, and fixes the faults that are found.

↓

Reviewer agent

A separate agent

Reads the work fresh and is asked to find faults, ranked blocker, major or minor. The work goes over; faults come back.

↓

Second builder

Another AI system

Working separately, it audits the results independently, cell by cell.

The roles, simplified. No data is shown.
01 / BUILD

One builds

I agree a plan with the builder agent before anything is written. It writes the program and its tests, and fixes what the reviews find.

02 / CHALLENGE

One challenges

A separate agent reviews the work without relying on the builder’s account of it. It looks for faults and ranks them: blocker, major or minor.

03 / AUDIT

One audits

A second builder, on a different AI system, works apart from the first. It audits the results cell by cell.

In more detail. The reviewer works from the code and the files as written, runs small tests of its own and reports each finding with its severity. A blocker is anything that could harm an approved result or approve a wrong one; nothing runs on the real data while one is open. The builder and the independent auditor run on different AI systems, so the audit is a separate check on the builder’s work.

An explanatory reconstruction of how the work is shared. Roles are simplified.