matrix.
everything i think about ai. quotes, takes, bugs, model notes, prompt recipes, bets. things that don't fit on twitter, which is most of them.
essayMay 12, 2026 · 8 min · model: opus 4.7
the orchestrator problem: when one model speaks for the others.
claude killed a codex audit mid-thought. that's the inciting incident. the actual essay is about who gets to decide when a worker has 'contributed enough'.
◆quotes · pick a treatment
"codex is still thinking. i have enough from both consultations now."— claude code (opus 4.7), 10:14:58, mid-audit
◆recent · everything else
subagents are an org chart you can rewrite at runtime.
managing six agents in parallel is just managing six junior engineers — except you can change their reporting line mid-task. that's the unlock.
today's pricing announcement is a compute tax dressed as a launch.
you're not paying for capability anymore, you're paying for context length. that's a different bill structure than the early days.
is it just me, or does CC rewind soft-lock the composer?
rewind to a specific checkpoint, try to type, nothing takes. only ctrl+c → resume fixes it. anyone else?
same tool name across two MCP servers silently picks the last-loaded one.
repro in 3 lines. workaround: alias one of them via the config. it's a foot-gun for anyone running > 1 server.
skills loaded from ~/.claude/skills don't appear in /help when a project-local skill has the same name.
project-local wins, but /help doesn't list the global one even as 'shadowed'. confusing for the cohort i was teaching last week.
sonnet 4.6 is the right default for review. 4.7 still over-explains.
5 of my last 7 code reviews: 4.6 catches more real issues, 4.7 generates more 'plausible-sounding but wrong' suggestions. flipping the default.
opus 4.7 is the right planner. don't let it touch the keyboard.
plan with opus, execute with sonnet, audit with codex/gemini. opus's plan quality is consistently better; its code is hit-or-miss.
gpt-5.5 finally says 'i'm not sure' without prompt engineering.
old gpt-5 confidently hallucinated. 5.5 hedges when it should hedge. small thing, big shift.
the matrix is what i think about ai when i'm not posting to twitter. take with a grain of salt — sometimes models change my mind by next week.