Sub-agents are not aware of their own time limits

[ammario]

I'm not sure how other ADEs handle this, but I've noticed that sub-agents (esp GPT-5 class ones) often time out. I wonder if we inserted a time_left_seconds: <number> in every tool call response when there's an overall task timeout if that would give the sub-agent sufficient information to better use available time.

thoughts @ThomasK33?

[ThomasK33]

I don't think that makes a lot of sense, since the concept of time is foreign to LLMs.

There are two approaches (non-exclusive) for us to improve the current status quo:

• add a maximum turn counter and inform the model at the beginning of the session how many turns are available (since the number of tool calls is easier to estimate than whether make test will take 3 or 5 minutes)
• add a "soft termination" by which the sub-agent gets interrupted, and a new message is sent informing it that it has reached its allotted resources and must respond with an agent_report with its current progress, even if incomplete.

The maximum amount of time still has to stay as a safeguard to avoid models from spinning endlessly (I had that with GPT-5.2 twice and multiple times with Gemini models).

So I'd rather introduce some sensible upper bounds and a soft landing here.

[ThomasK33]

On a side note, for the explore sub-agent, I'd recommend using one of the smaller models like Haiku or Gemini 3 Flash rather than the GPT-5 family. I've had better experiences with them in the past for this kind of exploratory/indexing kind of tasks.