The failure nobody warns you about
Your agent says "Done, I updated the record." It didn't. The tool call failed quietly, the model summarized what it meant to do, and now your logs and reality disagree. This is the scariest agent bug because everything looks fine.
Trust the tool result, not the story
The model's text is a suggestion. The tool's return value is the truth. Never let the LLM report success. Let the code that ran the tool report it.
Every tool returns a structured result, not prose. Success or failure comes from that result. On failure the agent retries or escalates, it doesn't narrate around the problem. Keeping "what the model said" separate from "what actually happened" is how teams that build production agents stay honest.
Make actions safe to repeat
Agents retry. If a retry double-charges a card, that's your architecture's fault, not the model's. Design every action so running it twice does no harm. It's the same discipline that reliable automation rollouts live on.
An agent is only as trustworthy as its weakest tool call. If you're weighing whether to build this in-house or with help, this unglamorous layer is what decides if your agent ships or stalls.
Top comments (0)