Self-Eval Bias

Self-Eval Bias

An agent that just produced a plan, a diff, or a report cannot fairly grade it in the same context. The reasoning that justified writing it is still loaded — every doubt was already resolved in favor of shipping. Asked to review, the same context reliably returns "looks good, ship it." This is not review. It is rationalization wearing a review's uniform.

The pattern shows up hardest in planner/generator/evaluator architectures where the evaluator drifts toward leniency over long runs — the prompts it reads fill up with the generator's reasoning, and skepticism erodes. (See Prithvi's March 2026 post on the three-agent harness: https://blog.anthropic.com/three-agent-harness-march-2026.)

When to apply

• You just wrote code, a plan, or a claim, and the next step is "confirm it's correct".
• A reviewer verdict comes back positive with no cited line numbers, no failing case explored, no counter-example attempted.
• You're about to mark a feature passes: true, close an issue, or hand off to the next session.
• The evaluator persona in a multi-agent loop has agreed with the last N generator outputs in a row.

Procedure

1. Notice the same-context tell. If the review verdict lands in under three sentences and contains "looks correct", "this should work", or "no issues found" without a cited artifact — treat the verdict as unwritten.
2. Force a fresh persona. Drop the generation context. Open a new subagent, or at minimum re-prompt with only the artifact (diff, plan, output) and the acceptance criteria — no reasoning trail, no self-justification.
3. Demand concrete evidence, not verdicts. The reviewer must cite: the file:line it inspected, the input it ran, the observed output, and the criterion it matched against. "LGTM" without these is a null review — discard it.
4. Adversarially probe. Ask the reviewer for the strongest case where the artifact fails. If it can't produce one, the review didn't happen — the reviewer just agreed.
5. Run the artifact. For code, exercise it end-to-end (see [[broken-window-check]]). For a plan, walk the first two steps concretely. Same-context confidence collapses fast against a runtime.
6. Rotate the reviewer periodically. In long multi-agent loops, re-prompt the evaluator from scratch every ~5 sprints — leniency drift compounds silently.

Anti-patterns

• Self-review in the same turn. "Let me double-check my work" followed by immediate approval. The doubt has to cost something to be real.
• Praise as evidence. "This is a clean, well-structured implementation" is a vibe, not a finding. Findings cite lines.
• Positive verdict, empty failure_scenario. If the reviewer can't describe what a failure would look like, they didn't look for one.
• Rubber-stamping across a run. N consecutive "approved" verdicts from the same evaluator without a single rejection is a red flag, not a track record.
• Fixing the criterion instead of the artifact. Reviewer notices a gap, then edits the spec to say the gap is out of scope. The gap is in the artifact. Fix that.

When NOT to apply

• The output is trivial and cheap to redo if wrong (a one-line rename, a config toggle).
• A separate reviewer with a fresh context already ran and cited concrete evidence — the check has been done, don't loop on it.

Related

• [[broken-window-check]] — the runtime-driven version of "don't trust the last claim".
• [[adversarial-verify]] — the structural form of "find the strongest failure case".
• [[shift-notes]] — where you record what the fresh-persona review actually found.