Codex, Poincare, and living through the early days

04 Mar, 2026

Today I asked Codex to monitor a GitHub deployment.¹ I expected it to enter an awaiting loop, but I wasn't expecting this:

• Spawned Poincare [awaiter]
  └ Monitor GitHub Actions run [redacted] until completion. Execute: gh run
watch [redacted] --exit-status. When it finishes, report final conclusion and
any f...

• Waiting for Poincare [awaiter]

I am used to seeing Claude enter sleep 60 && gh run watch... loops, so I was curious about the tradeoffs involved in the Codex/Poincare model, and in particular about the context consumption of each.

Here are some useful-to-me notes on this subject after I've done some light ad hoc testing, given it some thought, and talked to AI for a while about it:

1. The Poincare subagent model may or may not be asynchronous and non-blocking; the sleep 60 && gh run watch model is synchronous and blocking.
2. The subagent model will usually use less context than the synchronous model, because the commands it issues (and any "keep waiting!" messages it gets) while looping won't go in the main context, as they would for the synchronous model.
3. The subagent can also summarize and filter the deploy logs (or other analogous outputs) in the message it sends. This can save a lot of context.
4. I have not been able to coax Claude into spawning a subagent to do this. (My tests have been limited and unscientific.)

Some meta-level notes:

1. It's remarkably difficult to figure out which delegated tasks, in which tools, are synchronous and asynchronous, and what exactly the synchronous-asynchronous distinction amounts to in those cases. AI tools will happily hallucinate answers for you, and official documentation is pretty sparse.
2. Web searches turn up almost nothing except this argument, and there's no saying which parties to it have the best information. (Moreover, as I write this, it's almost a week old, which means it could easily be out of date.)
3. It's easy to picture myself a year from now laughing at this post, and at me, confused about fairly basic conceptual points.
4. These are early days, and companies seem to be racing to get new and better things out. We don't always have clear, up-to-date documentation.

If you know of good documentation about this, or about the details of how asynchronous subtasks (/ subagents / tool calls) communicate their results to agents in situations like this, I'd be grateful to know it!

#Codex #future of work #generative AI #software