How do I catch an agent that claims a task is done when it isn't?

Grade the state a session actually left behind, not the agent’s closing message. A completion grader compares an end state, a payment reference, a ticket status, a calendar entry, against what the user asked for, so the words “done” or “confirmed” in a transcript carry no weight by themselves.

ReliabilityBench shows why the distinction matters. Its travel domain, built on tools like search_flights, hold_flight, and confirm_booking, scores 87.5% pass@1, and the paper’s own failure analysis names the gap directly: agents fail to provide payment information for booking confirmation, a missing-context problem, not a broken call. Picture what that looks like in a trace: every step up to confirm_booking can return a valid response, so a checker that only reads step outcomes passes the run, even though confirm_booking never fires and no reservation exists.

The fix is to check the system the agent was supposed to change, not the trace of calls that got it there. If the booking system shows no reservation, the run failed, whatever the last message claimed.

sources

keep reading

More on this.

Are most agent errors caused by the agent's own logic?No. In Datadog's 2026 report, 5% of LLM call spans errored in February 2026, and 60% of those were exceeded rate limits, not the agent's reasoning.answer →Can content an agent retrieves override its own instructions?Yes. Text inside a webpage, document, or tool result an agent reads can contain instructions the model follows, a failure called indirect prompt injection.answer →How do multi-agent systems actually fail?Across three root causes: 41.77% specification issues, 36.94% inter-agent misalignment, and 21.30% task verification failures, per the MAST taxonomy's 1,600+ annotated traces.answer →Everything on failure modesThe definition, and every question answered under failure modes.concept →

Send us the traces you already emit.

book a demo