Your agent reports "Done! The record has been updated," but the record was never updated. How do you detect and prevent this?

Root causes to check in traces

1. No tool call at all: the model "imagined" doing it. Common when instructions push for helpfulness, or when the tool wasn't available.
2. Tool error ignored: the result said {"status":"error"} or was empty, and the model glossed over it.
3. Ambiguous success: the tool returned 200 OK but the write didn't persist (async queue, validation dropped it).
4. Truncated context: the error was cut off during compaction.

Prevention

1. Explicit tool results: {"updated": true, "record_id": "...", "new_status": "closed"} or a clear error, never an empty body.
2. Post-condition checks in code: after update_record, read it back. The tool returns success only if the post-condition holds.
3. Ground the final response: the user-facing confirmation is generated from the verified tool result (template or structured), not free-form model text.
4. Action ledger: the UI shows actions actually performed, from logs. The model's narration is secondary.
5. Prompt: "Only report actions as completed if the tool confirmed success; otherwise say what failed."
6. Evals: compare claimed outcomes with environment state; track a "false success rate" metric. This is one of the most damaging failure modes for user trust.

Related

• Previous: Q46. How do you keep an agent effective on long-horizon tasks that run for hours (e.g. a large migration)?
• Next: Q48. Design a multi-tenant platform where customers can deploy their own agents (code execution, tools, memory).
• Langfuse
• DeepEval