← sources

What information do you actually want when an AI agent fails?

source: reddit (reddit:AI_Agents)published 2026-10-04 17:55ingested 2026-10-05 02:36

retained text

**I have been away for some days (busy with something new coming up!!!)  but I need help on this**

I ’ve been thinking about agent evaluation beyond just measuring whether a task succeeded or failed.

For example, if an agent fails 3 out of 10 tasks, I'd want to know if I chose the wrong tool or made a bad decision early in the workflow and I don't think receiving a response like "70% success rate" tells **what is going wrong**. 

For people building or evaluating agents, what information do you actually find most useful when debugging a failure?

Do you mainly look at the final output, the tool calls, the execution trace, or something else?

document fetched 2026-10-05 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wxlyyk/what_information_do_you_actually_want_when_an_ai