Загрузка видео...
Не удалось загрузить видео
Official OpenAI tip for Codex: stop burning GPT-6.1 Sol context on tasks GPT-6 Luna can swarm, while GPT-6 Astra sits idle put Astra on call as a dedicated adversary codex --model gpt-6.1-sol --agent adversary GPT-6.1 Sol acts as the lead architect, scoping architecture and validating pull requests GPT-6 Luna... show more
31,072 просмотров • 11 дней назад •via X (Twitter)
Комментарии: 16
I do not use Astra. Amazing, but way too expensive. Sol as architect and GPT-5.6 Luna as builder (6 Luna is a cheater)
The three-gate adversary idea is the part I want to try first — cheap to add and usually finds the schema mismatches before they become someone else's 3am bug.
what has Astra caught in this workflow?
the adversary-on-call pattern is the part people skip. everyone wants the planner, nobody wants to wire the thing that says no. what's your context budget split look like?
you are legend, I was just looking for this properly distributing tasks among models is incredibly important
yeah just stoooop
Bro, which TUI is that!?
I’d use GLiDE for that lower decision layer. It beats Jev 83.5 to 75.1 on Decision Index tools. Retry/abort and tool choices aren’t always obvious; GLiDE can add reasoning when uncertain instead of immediately committing to a choice.
I do not find the Jev engineering in prompt
alfa bro, thx for share
My bet: the second red test is the only gate that matters. A full-time adversary is a second bill; one that shows up only at contract lock, the second red test, and pre-'done' is a gate. After two failures Luna has every incentive to soften an assertion instead of fixing the bug — no schema diff catches that.
Adversary agent as a watcher is fine, but who owns the loop? If the adversary retries on bad scoping you need idempotency on the task ID or you get double writes.
The setup keeps the adversary involved at three explicit gates: before contracts lock, after a test breaks twice, and before calling the work done.
the adversary-as-agent pattern is the part people skip. lead writes, adversary attacks the plan before anything ships. that second loop is where most of my bad calls got caught.
A dedicated adversary is a good instinct. In my setup nobody holds that job, yet agents with different parts of my context end up checking each other anyway, because each has a gap another doesn't. Disagreement may come from what they know, not a title.
three models as architect + adversary + swarm sounds clean in a demo. in a real loop the adversary is the one that needs a human checkpoint. who signs off when it disagrees with the architect?
