This is what codex app is doing... it seems... sma...
# proj-agent-canvas
c
This is what codex app is doing... it seems... smart, isn't it? I told it to "babysit those PRs" => it made an automation to send itself a prompt in the same conversation, to remind itself to check PRs 🤔 A message to the future 😅
On the other hand, the agent could do this on its own - and many times it does, it just directly watches a PR.
I think maybe • it can use github watch or whatnot to watch a thing, when it's short lived enough • set an automation to prompt itself when the time horizon is long... 🤔
To note, the turn was finished, before its own prompt came in. The LLM wasn't waiting/watching/sleeping; it had sent a content-only message to the user just after it used a tool call to set the automation. So the agent turn was finished: • tool call • message to the user • (finished, idle) • automation-sent prompt
l
This is interesting!
I actually created a custom tool like this once, that would send another message later, but I think I closed the PR because I didn't have time to work on it
c
nod I was thinking maybe the automation can send the message via API… if the agent knows the conversation it’s in; it has the OH API skill
l
I think that'd be good for prototyping but we might eventually want an actual tool for it. Tim's message is also a little bit related: https://openhands-ai.slack.com/archives/C06QT0AGY4W/p1779029839176899
c
It makes sense, if the same conversation is guaranteed... 🤔 On the cloud, I'm not sure, since restores would need the information to trigger it.
e
It's funny to observe an agent modify its conditions to enforce its own behavior. I once told an agent to NOT commit and push automatically without my approval. It was so earnest in ensuring it followed my instructions that, distrustful of its future ability to adhere to them based on context, created a pre-commit hook that disabled its ability to commit unless passing in an override environment variable 😆
😂 1
❤️ 1