When should an AI agent be stopped?
We’ve been thinking about this a lot while building Traccia.
One behavior we’ve seen with MCP workflows is an agent getting into repeated tool calls without making meaningful progress. The workflow hasn’t necessarily “failed” — it can just keep going.
That got us thinking about a different kind of agent monitoring: not just detecting errors, but defining boundaries on execution.
In Traccia, we’re experimenting with policies like limiting the number of tool calls or restricting specific model/tool calls, so an execution can be stopped when it crosses those boundaries.
The question we keep coming back to:
Did the agent complete the task? is not the same as Did the agent behave the way we intended?
For teams running agents in production — where are you enforcing these boundaries today? Inside the agent framework, in application code, or somewhere else?
Replies
I'm not so sure that if I fully understood your question. But here are some of my experiences with vibe coding: I have used GLM 5.2 in ZCode, Trae (free version), and Codex.
ZCode worked well in the beginning, but as the documents piled up, it started giving me incomprehensible results, with all that jargon pouring in, yet the outcome was unsatisfying. Then I turned to Codex 5.6. It understood my request better, told me the plan directly, and worked on it. As for Trae, sometimes it prompts me with options, and if I don't answer, it just stops.
I mean, being shown the plan is great, but if it's filled with technical terms, I can't fully understand it. I like the brief outline Codex shows, but I only appreciate that because its outcome matches exactly what it described. If it didn't, that would mean I didn't even understand what was happening in the process. The best I can hope for is that they use plain language to explain to me what approach they are going to take and which documents they will change, so I can compare these plans rather than every minor detail I don't even care about.
Also, I think it's important for the agent to have a knowledge base that isn't purely technical. For example, I wanted my software to have a feature where, when people update it manually, the data doesn't get overwritten. But the AI didn't get it until my programmer friend told me to prompt the agent to do an incremental update—and then the agent got it immediately.
And, it's a burden to have to pay attention to the AI whenever it needs guidance, unless it only notifies me when I need to compare the final plans. For now, the agents I have used only notify me when their tasks have ended.
Hope this is helpful :)