# How do you teach an agent when to stop trying? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/cb074f9474ce63cf

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-24 09:13 ingested 2026-10-25 02:35

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xe3zp3/how_do_you_teach_an_agent_when_to_stop_trying/)

## retained text



```
One thing I keep running into with agents: they can keep doing reasonable-looking work long after it stops being useful.



Another review. Another small fix. Another follow-up. Each action makes sense on its own, but the overall task isn’t really moving forward.



Say an agent is trying to get a refund by email. After several attempts, the useful next step might be to suggest calling support, rather than draft another email. Or before writing custom code, check whether an existing SDK already does the job.



And sometimes the task is already done. The result has been verified. There’s no reason to keep looking for things to improve.



I want agents to get better at deciding whether to continue, change approach or stop. A limit on iterations prevents an endless loop, but it doesn’t tell the agent which of those choices makes sense.



I’ve tried making the desired outcome, constraints and stopping criteria explicit, and putting recurring lessons into persistent instructions. These are observations from using agents, not controlled experiments.



The question I want the agent to ask is: is the next action worth the time, cost and risk, given what we already know?



How are you handling this? Have you found something that improves that judgment, rather than just cutting the agent off after a fixed number of steps?
```



document fetched 2026-10-25 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xe3zp3/how_do_you_teach_an_agent_when_to_stop_trying
