An agent PR that only says 'fixed' makes the reviewer redo the diagnosis

An agent finishes a task, reports 'done, tests pass,' and opens a pull request. The reviewer opens it and finds no reproduction, no before state, no note on which check actually changed. They now face a choice: trust the claim, or re-derive the failure from scratch. A green suite proves the current checks pass; it rarely proves the specific bug that broke is closed.

A green suite does not prove the bug is closed

The existing tests cover what the suite already knew how to check. A fix that passes them can still have patched a nearby symptom, changed an input so the old failure no longer triggers, or relaxed the check instead of addressing the cause. The reviewer cannot tell which happened from a one-line summary, because the summary describes intent rather than evidence.

This is where review time disappears. Instead of reading a diff, the reviewer rebuilds the failing scenario, reruns the command, and guesses at the edge the agent meant to cover.

The evidence a reviewer actually needs

• The reproduction. The exact command that failed before, the expected result, and the actual result. Without it, the reviewer cannot tell whether the agent fixed the reported bug or a neighboring symptom.
• The before/after check. The test or check that failed and now passes, and whether it was added or already existed. A new test written against the fix proves coverage; an old test whose behavior changed proves little on its own.
• The touched scope. Which files changed and why, plus anything deliberately left out. A small diff that reaches three unrelated modules signals the agent drifted from the task.
• Uncovered assumptions. Inputs, states, and edge cases the fix did not exercise, stated plainly. Silence here reads as 'all cases covered,' which is almost never true.
• Revert and risk. How the change rolls back and what side effects the reviewer should watch for after merge.

Give the model a record, not a narrative

Agents default to a summary because summaries are what they are trained to produce. A fixed PR template with blank fields for reproduction, before/after check, scope, and assumptions turns the output into a record the reviewer can check line by line. Add 'repro command present' and 'failing check named' to the agent's own done-checklist; it stops marking done when those fields are empty.

The template does the discipline for you. Without it, the model writes a confident paragraph; with it, it fills facts that survive review.

The cheap reviewer move

Before accepting 'done,' ask the agent to paste the failing log line and the command that reproduced it. If it cannot, the work was verified against memory of the bug rather than the bug itself. That one exchange catches most cosmetic fixes before they reach review, and it costs the agent one short turn.

The point of the PR body is not to impress the reviewer; it is to let someone verify the claim in minutes instead of repeating the diagnosis. A summary asks for trust. Evidence asks for a check.