Troubleshooting

Your Agent Said It Did Something… But Nothing Changed?

A fast way to diagnose when an agent sounds confident but the action never actually happened.

This is a trust-breaker that shows up often: the agent says “done,” but no file was created, no message was sent, or no change appears in your workspace.

Important: polished language is not proof of execution. Always verify outcomes, not wording.

2-minute reality check

Ask the agent to show specific evidence (path, command output, commit hash, or screenshot).
Confirm the result yourself in the target place (file exists, message posted, task completed).
If there is no evidence, rerun with a deterministic prompt (template below).

Common reasons this happens

Planning language confusion: “I’ll do X” was interpreted as completion.
Wrong workspace/path: action happened somewhere else.
Permission/approval stop: the tool call was blocked and never executed.
Context drift: the agent responded from chat context without re-checking current files.
Surface mismatch: output landed in a different channel/thread/session.

High-friction case: “Approval Required” looked like completion

One of the most common false-"done" scenarios is an approval pause. The agent may describe the intended action, but execution is waiting for you to approve the exact command.

Check whether the run shows an Approval Required step.
Approve the exact full command (including &&, pipes, semicolons, or multiline script content).
After approval, ask for fresh execution evidence (not the pre-approval plan).

Gotcha: approving a partial command often causes repeat loops. Approve the complete command string exactly as shown, then verify output again.

Use this “show your work” prompt

Do the task now, then return ONLY: 1) What changed 2) Exact evidence (file path, command output, or link) 3) A one-line verification step I can run

For file edits: require before/after proof

Request the exact file path.
Ask for the changed section (or line range) before and after.
Ask for one verification command, e.g. grep or git diff -- file.

For external actions: require destination proof

GitHub: PR/commit URL.
Discord: channel/thread reference and message confirmation.
Cron/reminders: job ID and schedule details.

Best practice: ask for “done + evidence + verification” every time you care about reliability. This reduces support loops dramatically.

Escalation packet for #help

Execution mismatch report Task requested: What the agent claimed: Expected result location: Actual result: Evidence provided by agent (if any): Your verification step + result: