25 Opus threads on under 20% of a $200 plan
Theo kept between 8 and 25 Opus 5.5 threads running all day and still burned less than a fifth of a $200 Claude subscription. The week's theme was throughput: more parallel agents, cheaper subagents, and fewer tokens spent arguing with your own code reviewer.
The big disagreement
poteto calls code review a token waste, and the week mostly agreed
poteto said he now considers adversarial reviews mostly a waste of tokens, preferring empirical plus formal verification, and has cut the interrogate skill from pstack. Thorsten Ball endorsed the "irrefutable proof that this works" framing with a clip showing one. Matt Pocock arrived from the other direction, arguing readability now means verifiability and that cyclomatic complexity and function size do not matter to agents.
the smartest models will need fewer and fewer tokens, they won't need to verify their work or check it for bugs, they will just be right
Receipts
Theo ran 8 to 25 Opus threads a day on under 20% of a $200 sub
Theo says he had between 8 and 25 Opus 5.5 threads going most of the day and still used less than 20% of a $200 Claude plan, while his Codex subs got hammered because he uses Sol for reviews. He also claims T3 Code users more than tripled their token throughput after switching apps.
The amount of value you can get out of the $200 Claude plan rn is insane.
dax's DeepSeek Flash subagents held up, and he posted the rules
After a week of running all subagents on DeepSeek Flash with Opus 5.5 only on the primary agent, dax says it is working great and listed what makes it work: long running sessions per project, strict scope on every task, proof of work instead of a summary, and a one step review of the diff. He had tried the same split before and it was not as good.
ask for proof of work instead of a summary
How they actually work
Pocock's prompt: how do we make this change fully verifiable?
Matt Pocock's prompt of the day is one line, and he says it makes the agent bend over backwards: signing up for real services, clicking through the frontend, quality through the roof. David K Piano posted the same instinct more bluntly, that telling agents what you want beats telling them what to do. Kent C. Dodds added that you should be talking with your agent, not at it.
You should be talking with your agent, not at your agent.
Open tabs
Willison got the GPT-6 Astra weights trick out of a laptop model
Simon Willison tried the buzzy weight-interpretation paper with Qwen 3.8 27B on his laptop and says it reached the same result after 45 plus minutes.
Frequently asked questions
Why are developers moving away from code review?
The cost has shifted because formal verification and empirical testing are now cheap to generate, while detailed code review feedback has become expensive in comparison. This reverses the old tradeoff where testing was costly and review was free.
What's the bottleneck for AI coding now instead of subscription limits?
The real constraint has moved from hitting subscription limits to how many parallel agent threads one person can actually supervise and manage at the same time.
How do you make cheap models work well for coding tasks?
The four key rules are: keep long running sessions per project, set strict scope on every task, require proof of work instead of summaries, and do a one step review of the diff. The orchestrator prompt does the work, not just the model choice.
What prompt style gets better results from coding agents?
Instead of telling agents what to do, state the outcome you want and the standard of proof needed. Agents perform better when given the goal and verification requirements rather than step-by-step instructions.
Can weight interpretation techniques run on consumer hardware?
Yes, the technique was demonstrated working on a laptop model after extended processing time, achieving similar results to larger implementations.