FDE Times: when your agent is wrong for the first time — contain, measure, then speak

FDE Times published a 6 November guide on what to do when an agent you deployed is wrong in front of users for the first time. The piece argues that “the model hallucinated” is not a defence: Moffatt v. Air Canada treated the chatbot as part of the airline’s website, and Cursor’s support bot “Sam” inventing a one-device subscription rule shows how quickly a fabricated policy spreads.

In the first hour the jobs are contain, measure, then speak. Route only the broken topic (for example returns) to humans so the rest of the agent keeps running; search logs for every conversation that stated the wrong figure; then send the customer’s owner a short note with what happened, the correct fact, the affected count, and a redress proposal that needs their approval. Avoid blaming the model, a person, or declaring “all fixed” before you have numbers.

Afterwards, fix systems not people: remove stale RAG docs with effective dates, add eval cases for each policy figure, and alert when the agent states a number that disagrees with the source of truth. Anthropic’s September 2025 quality postmortem is cited for continuous production evals; reopen the topic only after those cases pass and share a one-page postmortem plus a 72-hour faithfulness chart with the customer. For FDEs, that incident story — how fast you contained it and what the new evals caught — is also interview-ready material.