Struggling on Plus? I got Astra to act like a proper Codex manager instead of doing all the work itself

I’ve been experimenting with Codex multi-agent workflows because I was getting tired of manually doing this:

Astra plans something, I ask it for a handoff, open a new Sol session, paste the handoff, let Sol work, copy the result back to Astra, ask Astra to review it, then repeat.

It worked, but I was basically acting as the message bus between models.

So I built a project-specific AGENTS.md that makes Astra handle the orchestration itself.

The main idea is that Astra should behave more like a technical lead than the smartest coder in the room.

The instructions tell it to use Luna for cheap repo exploration and mechanical work, Terra for normal implementation, Sol for harder coding/debugging/review, and Astra itself for architecture, decomposition, ambiguity, escalation and final judgment.

I also separated model choice from reasoning effort. So a task doesn’t automatically get “High” just because it’s going to Sol, and Astra doesn’t waste high-effort reasoning on simple coordination turns.

One thing I was especially worried about was token usage. I didn’t want Astra constantly polling workers and repeatedly burning context while another model was doing the actual work. So I added explicit rules telling it to delegate, wait efficiently, and only wake up when there’s something meaningful to review or correct.

I also told it to keep child contexts as small as possible instead of blindly passing the full parent session into every worker.

The latest refinement is probably the most useful part.

At the end of every meaningful assignment, Astra now has to report:

- which agents/models it used

- what reasoning effort it requested

- whether that routing was actually confirmed or only requested

- how effective the delegation was

- whether it used too many agents, too few, or roughly the right amount

- what it learned from the assignment and how it should adjust future delegation

So if Terra handled something easily and Sol review added nothing, Astra should explicitly say that the next similar task probably doesn’t need a Sol reviewer.

If Terra struggled and it had to escalate to Sol, it should learn that similar work may deserve Sol from the start next time.

I also added context-budget awareness.

At the end of a task, Astra has to assess the current session and tell me whether it’s still sensible to keep working in the same context, or whether the session is getting too crowded and it would be better to prepare a handoff and start fresh.

That was another thing I used to manage manually, usually after the session had already started getting bloated.

I’m keeping all of this project-specific for now rather than putting it in my global AGENTS.md , because I want to see how well it behaves across different types of work before making it my default.

The Codex desktop UI is still a bit frustrating for monitoring this. It shows spawned subagents and whether they’re working or done, but it doesn’t currently give a very clear live view of the actual model and reasoning effort being used. That’s partly why I added the explicit reporting requirements.