Agents are often asked to review their own work. The pattern feels efficient:
generate an answer, critique it, revise it, and stop when the same model says it
is good enough.

Self-review is useful for iteration. It is weak as the final authority.

The agent shares the assumptions, blind spots, context, and incentives that
produced the first answer. A second pass can catch obvious mistakes, but it
does not create an independent reason to trust the result.

## Keep the useful part of self-review

Let the agent inspect its draft. Ask it to find missing evidence, test edge
cases, and explain uncertainty. This can improve quality before a final check
and reduce avoidable review work.

Then separate the decision that has consequences:

- the coding agent may run tests, but a protected job decides whether the pull
  request can merge;
- the support agent may draft a refund, but a policy service decides whether
  the amount is permitted;
- the research agent may summarize sources, but citation checks use the source
  documents rather than the agent's confidence;
- the publishing agent may prepare a post, but a person approves claims that
  affect reputation or legal exposure.

The goal is not to add a human to every step. It is to give the final decision
an authority that is not controlled by the proposal being judged.

## Match independence to the risk

Not every agent action needs the same level of separation. A private draft can
use lightweight checks. A production deployment, payment, account change, or
public claim needs stronger evidence and a clear handoff.

A practical scale is:

1. **Draft:** the agent acts in a private workspace and explains its work.
2. **Checked:** automated tests or policies inspect the result.
3. **Compared:** the result is measured against a saved baseline or held-out
   examples.
4. **Approved:** a separate service or person authorizes the consequential
   action.

Move up the scale as the blast radius grows. Make the level visible so users
know whether they are looking at a suggestion, a checked result, or an approved
action.

## The design rule

The agent may help produce evidence. It may not be the only authority deciding
what that evidence means.

That one rule makes agent workflows easier to debug, easier to explain, and
safer to expand over time. It also lets teams change models without rebuilding
the meaning of success around every new model.

## Continue

- [Read the starting guide](/notes/control-loop/a-passing-test-is-not-a-good-result)
- [Learn the two-part architecture](/reference/two-plane-loop)
- [See how Cursor performs against the same questions](/notes/teardown/scoring-cursor)