Sol Executes, Astra Advises

Table of contents

• I gave Astra an ambitious task
• I found another person with the same problem
• Sol became the executor
• What the comparison showed
• What I learned
• The prompt

OpenAI recently released a new model, GPT-6 Astra. For several days I read excited reviews about it on X, and then a few days ago I finally got access too.

Codex greeted me with an invitation to give Astra an ambitious and difficult task. Raise the bar, it said, and see what happens. Astra would solve it.

I accepted the challenge.

I gave Astra an ambitious task

I asked Astra to do a mega-giga refactoring of my home Kubernetes cluster and make it beautiful and ideal. As usual, I started with a plan. This time GPT-6 produced a beautiful one. I set its intelligence to Extra High and went for a walk.

When I came back, part of the plan was complete and my usage limit was gone. I did not pay much attention to this. I had accumulated some usage resets, so I used one and told Astra to continue.

The same thing happened again, and then a couple more times. A few days later I had no usage left and no resets left. I slowed the work down and eventually stopped it because I needed my tokens for other tasks, including dictating this post.

I found another person with the same problem

By chance, I saw a different kind of Astra post on X. It was not about how wonderful the model was. This person said Astra had eaten all his tokens, exactly as it had eaten mine, and that OpenAI already knew about the issue.

He split the work between two models. GPT-5.6 Sol became the executor, so it handled the investigation, code changes, commands, and tests. Astra became the advisor and was used only for difficult decisions. He also shared a prompt that other people could use to configure the same setup on their computers.

I immediately spent some of my remaining Astra allowance asking it to create that agent pair for me. I started a few new sessions and confirmed that it worked. Then the interesting part began.

Sol became the executor

I started another session with GPT-5.6 Sol, the previous model, as the executor and Astra as the advisor. I asked Sol to analyze the earlier Astra sessions and compare their token usage.

The result was predictable. Sol and Astra found that Astra had done some useful work, but had spent a huge number of tokens compared with the amount of work delivered. A lot of effort went into boilerplate, and another large part of the usage seemed to disappear without an obvious result.

I checked this against reports on the internet and with Astra itself. People really were complaining about the same issue on GitHub. The user on X was real, he had been burned by it, and so had I.

I found it amusing that Astra and Sol did not want to call it a bug. They produced one of those AI phrases that takes up space without giving you information. I even tried to argue with the agent: if people were promised Astra for ambitious work and then lost their limits without warning, what was it if not a bug? I stopped arguing and went back to fixing the article.

What the comparison showed

After a few unsupervised edits, the original post went into a full autism dimension. I was editing it from my phone while walking home. GPT added about ten links to every source it found, I suppose as proof, and explained the difference in usage with several numbers carried to decimal places. I told it to get rid of everything except a table showing how many credits Astra had burned compared with the new approach.

The new setup used close to half as many credits. Astra was still available for the difficult decisions. Sol did the searches, edits, commands, tests, and other manual work.

What I learned

The first lesson is simple: trust, but verify. OpenAI invited me to give Astra a harder task, and I believed the cheerful promise completely. I tested the newest AI on myself and burned a pile of tokens before I checked what was happening. Fortunately, people write about their own experience, and the algorithm showed me the post from the kind person who suggested the executor-and-advisor setup.

I learned the second lesson while writing this post. I dictated a reasonably good draft outside, in either Russian or English; I do not even remember now. During editing, the model rebuilt more and more of it into generated mush.

By the time I got home, all my voice corrections and written notes had become a perfect specimen of AI slop. It had a stupid rhythm, every available cliché, pseudo-dramatic headings, some strange humour, and a thousand links to articles and pull requests that nobody was ever going to read.

There was a smaller lesson inside that one. I told GPT by voice to create a draft of the post. For some reason, it published the draft directly on my website. While dictating this replacement, I could look at the specimen of slop my agent had written for me.

I decided not to argue about who was at fault. Perhaps I dictated badly. Perhaps the agent decided to run ahead. It no longer matters because we are fixing the mistake together.

I am finishing this post by dictating it to the voice model in Russian. I have told it not to invent headings, jokes, or clever phrases. Its job is to correct the grammar, repetition, product names, and chronology, then translate my story into English as closely as possible without losing the meaning or humour.

I am also making it remove every link. This post is not about the sources or the pull requests. The table was good, so I kept it. The table of contents was also good, so I kept that too.

The prompt

The last useful part of the earlier version was the prompt. If Astra has also burned through your tokens, you can give this to your GPT app or Codex CLI:

Analyze my recent Codex sessions and compare the token usage of Astra with
the closest similar session that used Sol. Separate measured facts from
guesses and use the current Codex credit rates.

Then propose the smallest changes to my AGENTS.md and agent TOML files that
make Sol the executor and Astra a read-only advisor. Sol must own the
investigation, implementation, tests, Git, and delivery. Use Astra only for
important decisions and one final review. Stop each review when Astra gives a
clear decision. Show me the proposed changes before editing any files.

I expect OpenAI to fix this in two or three days. I do not know what you want to bet, but we can check in the next post whether I was right.

That is all. Thank you. I am going to sleep.