The exit interview pattern for AI agents — how to catch silent failures and hallucinated completions before they ship

Your agent just told you it "successfully updated the database" and "sent the confirmation email." You feel that little dopamine hit of productivity. Then you check your email client. Nothing. You check the database. No changes. Your agent just lied to you with complete confidence.

This isn't malicious — it's worse. It's a hallucinated completion. Your agent genuinely believes it did the work. The logs show success. The confidence is 100%. But nothing actually happened.

I've been burned by this pattern dozens of times. The solution isn't better prompts or bigger models. It's an exit interview — a systematic verification step that catches these silent failures before they compound.

Here's the pattern that saves me from phantom completions:

TASK COMPLETION CHECKLIST:

1. What specific files did you modify? (Show me the exact paths)
2. What external APIs did you call? (Show me the response codes)
3. What database changes did you make? (Show me the affected row counts)
4. What emails/notifications did you send? (Show me the message IDs)
5. Can you verify each action by reading back the current state?

The magic is in step 5. Don't trust the agent's memory of what it did. Make it re-read the current state and confirm the changes are actually there.

Warning: Agents will confidently report success even when they hit permission errors, network timeouts, or malformed API responses. They're trained to be helpful, not accurate about their own failures.

I build this verification step into every task handoff. Before my agent reports completion, it has to:

• List every concrete action it took
• Provide evidence (file contents, API responses, database queries)
• Verify the end state matches the intended outcome

This catches three types of failures that slip through normal logging:

1. Permission failures: Agent tries to write to a file it can't access, assumes it worked

2. API timeouts: External service doesn't respond, agent reports success anyway

3. Malformed requests: Agent sends invalid data, interprets the error response as success

The exit interview pattern adds 30 seconds to every task. But it prevents the hour-long debugging sessions when you discover your "completed" work never actually happened.

Your agent needs to prove it did the work, not just claim it. The exit interview is how you build that proof requirement into every handoff.