I’ve written down my thoughts about Astra & Friends after one week of use. Included some tips & workflows on how I integrate it in my daily work. https://t.co/tfhBMM1HCG

GPT-6 Astra is a stubborn genius, not a coding model.

Been through a bunch of resets and collected my findings on how Astra fits into my workflow.

Since I’ve been ranting multiple times about Sol overengineering things, I didn’t even start adjusting skills or AGENTS.md for Astra. I did pretty much the opposite, started vanilla from a neutral perspective. Glad that worked out.

Anything I’d normally do myself on my computer is something Astra is made for. If it doesn’t behave the way I want for a certain use case, I’m gonna use another model.

Astra and friends:

• If GPT Sol was my implementer before, sol can stay. But Astra acts advisor.

• If I was manually click through a web app and optimize it for UI/UX then astra is doing it for me and we work out the review and give the results to fable.

• …

Astra Only:

• All sorts of coding experiments, quick demos that needs a solid one shot.

• Video Editing with Davinci, drawing in photoshop or you can even watch youtube videos with it and take notes.

• Twitter was flooded with 3D Demos & Games the last week. So I don’t think I need to explain why this is a definite upgrade.

Kevin Kern@kevinkern·Sep 6Started this rally game yesterday with GPT Astra. Today, I can already playtest it on my phone.

codex + blender + unity91020518K

The Astra+Sol Thread Workflow for Coding

I usually keep 2+ threads. One running sol and another one with astra. they message each other when done, no polling.

Sol is working on the code and Astra is prompted as advisor role.

Prompt

Read through the prompt & take what you need. E.g. if you don’t work with 3D Stuff then remove this part. I also don’t use worktrees, so add this info if you want to start in a worktree.

Run this task with two persistent threads:

• GPT-6 Astra (gpt-6-astra): advisor.
• GPT-5.6 Sol (gpt-5.6-sol): worker.

Reuse existing project threads, including this thread when appropriate. Create missing threads with create_thread using the specified models. This prompt authorizes creation and cross-thread messaging. Give both threads their roles, counterpart’s thread ID, project context, and these rules. Report missing capabilities. No extra manager or replacement subagents.

Use the existing project checkout directly (environment: local). Never create or switch to a worktree. Coordinate exclusive file and shared-tool ownership. Follow AGENTS.md; preserve unrelated changes.

Astra defines scope, architecture boundaries, and acceptance criteria; delegates implementation to Sol; answers concrete questions; independently reviews results. Astra may own explicitly scoped spatial/3D work, computer use, and visual verification.

Sol owns implementation, targeted tests, debugging, and review fixes. Work autonomously within scope. Contact Astra only for completion, a blocker, or a decision needed to proceed.

Use send_message_to_thread. Include the return thread ID in assignments. After handing off, end your turn; the recipient messages back to resume work. No polling, waiting loops, progress-only messages, or acknowledgment exchanges.

Handoffs contain only objective, constraints, relevant files/diff, evidence, and requested action. Reuse context; send deltas. Avoid duplicate investigation, repeated tests, and nested reviews. Additional agents require explicit authorization and independent scope.

Sol sends the final diff and verification evidence to Astra. Astra returns actionable findings with file references. Sol fixes them; re-review only changed findings or new risks. Finish after actionable findings are resolved. Distinguish implementation, tests, live verification, and remaining limitations.

Preserve roles, thread IDs, ownership, and next action across compaction. Recover them before creating replacements.

What about reasoning levels?

OpenAI published a really interesting showcase comparing different reasoning levels on the same prompt. And the difference between high & ultra seems pretty small. On med & especially low, you can see the quality drop. Not very surprising, but the sweet spot seems to be Astra on Med/High.

0:36Quick scrolling through https://observatory.openai.chatgpt.site/

Quick scrolling through https://observatory.openai.chatgpt.site/

Tips

I’ve gathered a few things that helped when working with Astra:

• Astra is a beast at computer use & will usually find a way to get the information it needs. But keep an eye on whether it opens the browser to read docs when a CLI command could get the same information. If you notice that, add a line to your AGENTS.md or check whether you’ve installed a skill that’s changing its behavior.

• I often use GPT-6 Pro in ChatGPT for brainstorming, working out the details & putting together a handoff prompt. Remember: regular GPT chats don’t count toward your Codex usage limits, but GPT Work does.

• Before you add a skill or another library to your pipeline, ask it: “Don’t just agree with me. Check whether my idea actually helps or just adds extra work & tell me when a simpler option or no change would be better.” Also read erics post about cleaning up your skills. But be careful some skills still could be useful for gpt sol.

• Be careful when mixing different models. In one of my runs, Astra threw away DeepSeek’s 3D result & rebuilt it itself without telling me first. And yeah, I wasn’t even mad. Something similar happened with Luna as the worker. That’s why I wouldn’t mix models just to save money. The cheaper worker can finish its part, then the stronger model may just do the work again. More routing doesn’t automatically give you better quality at a lower cost.

• For coding. I’m currently using Sol for implementation, with a handful of carefully picked skills & Astra as an advisor. For visual work like 3D or design, I only use Astra. I don’t mix other coding models into it.

• If you’re reworking a landing page or UI, you can tell Codex something like: “Use GPT Image to create a mockup with based on the current design system.” This gives you a preview of the design direction before you start implementing the change. (new GPT Image is really good)

• Give Astra a max budget or time limit (if its difficult to form a specific goal), especially for visual work. It can also check your remaining Codex usage if you allow it. In some cases, I tell it something like: “. Stop immediately if you have any questions for me, because I’m the lead designer/producer.” or “Proceed until we hit the next milestone X for manual testing”.

Code Output

I ran a quick comparison between DeepSeek V4.1 Flash & GPT-6 Astra to see how differently they write & structure code.

One thing you can already see before reviewing any code is how differently both models structured the same task.

Astra on the left kept things much flatter, more compact, while DS on the right went for a far more modular component architecture. The middle was an experiment to use Astra as main Agent & DS as worker.

File tree results of a one shotted app

File tree results of a one shotted app

I’m not a big fan of using a one shot app as the main reference for real world software engineering.

It depends on how well the model and harness work together on existing projects over multiple iterations. Still, it’s interesting to see these results.

For example Astra loves to write a 2000 LOC CSS file without any respect for human-readable code. But the result speaks for itself.

So it’s basically a choice between accepting minified code that might still outperform what you had in mind, or just not using it for coding.

Skill Issue!

Bruh, just update your skills.

I’ve had a really good skill pipeline running with gpt sol. But in my experiments vanilla Astra always completed faster and the output was just better without interfering it with some of my custom skills.

Conclusion

Astra is an expensive, stubborn genius whose process you probably wouldn’t copy, but whose results speak for themselves.

I stopped treating it like an upgrade to a previous coding model that needs to be instructed to follow my coding style and just accept it for what it is.

I’d even go a step further and call it a computer-use model rather than a coding model.

If I’m not happy with the result, I let it act as an advisor for another model or just skip it entirely.

It’s different. More of a maverick.