I’ve only been using Astra since Friday (less than two days) but I already have more trust than any previous model that it can get the job done, whatever task I give it.
It feels like I’ve hired an extremely competent employee with expert-level skills in every digital domain. The limits feel increasingly like my own imagination and how clearly I can specify the goal I want to accomplish.
Astra has been a long time coming, and I wanted to revisit the Codex setup I’d carried over from the GPT-5.x era. Here are six configuration changes I’ve made that, in my limited time using it, have helped Astra shine and that you can make too.
Free preview: Tips 1 and 2. The installation commands, multi-agent setup, and explore-then-implement workflow tips are for paid Attention Heads subscribers.
1. Turn on persistent context
Astra’s experimental context-management feature lets it save notes across context windows and search earlier messages and tool calls. That means a long task does not have to depend entirely on one giant, fragile summary that loses information the longer the thread runs.
In your Codex config.toml, add (or ask Codex to add):
[features]
context_management.experimental_mode = true
In my experience, this makes Astra feel like it can work beyond its active context window. This makes a long-running task feel more coherent, even as it spans multiple context windows.
That is a much better memory model than asking an agent to keep re-summarizing its own life. I’ve only used it for a short time but I imagine it’s the sort of mechanism that gets stronger as you go along. The feature is experimental but OpenAI said it will become the default in the coming weeks.
2. Remove the context that no longer deserves to be there
Astra is better at following instructions. That is excellent news until you remember how many old instructions you have accumulated.
Skills you forgot you installed. An AGENTS.md written for a previous model. A prompt that made sense six months ago and now tells the agent to ask permission three times before it opens a file.
Run this context audit prompt in Codex before your next session with Astra:
Audit the skills and AGENTS.md files available to this project.
Identify redundant instructions, conflicts, stale model-specific guidance,
and unnecessary approval pauses. Preserve real safety and publishing gates.
Propose surgical edits; do not change anything until I approve them.The goal is not to make your agent less constrained. It is to make the constraints legible, current, and proportional. OpenAI specifically recommends auditing accessible skills and instruction files because Astra is sensitive to the guidance that is loaded into its context.
A smaller, clearer operating manual gives Astra more room to do the work you actually asked for and removes the bloat accumulated from GPT-5.x era prompts.
One example from my instruction cleanup
My GitHub Actions repair skill used to require a fresh approval even when I had already asked the agent to fix the failing checks. The audit replaced that blanket gate with a rule that distinguished a repair request from a request for diagnosis.
The revised rule still asks about material decisions or consequential actions outside the existing authorization. It removes a redundant pause while keeping the scope boundary. This is an actual change from my September 5 audit; I have not measured its effect on task speed or quality.
🔒 Four more ways to give Astra real agency
The next four are where the model becomes less like an excellent individual contributor and more like the person running a small technical team.
A practical companion for paid subscribers
I’ve put together an Astra Field Kit to accompany this piece: an interactive configuration builder, example configs, prompts, and a checker to help you catch mistakes before you use them. Paid subscribers get lifetime access to this edition of the kit.
