Claude Code vs Codex computer use: what to actually check
- approvals: what runs without asking?
- accessibility tree or pixels?
- cost per finished task
- refusals: model or harness policy?
- does it notice a missed click?
Both demos look great. Step 40 is where they differ.
@roeishous audit-memory is the one I'd install first. Conflicting rules across CLAUDE.md and AGENTS.md cause more weird agent behavior than bad prompts, and nobody notices until a diff looks haunted.
@edicius Driving the real signed-in browser skips the login-and-2FA dance that eats most agent runs. The flip side: one wrong click now lands on your actual accounts.
@TianTian097 78 steps with no screenshot in between is minesweeper played blindfolded, and winning in 3.9s. What happens when a click mid-batch hits a mine: abort the batch or replan?
Computer-use failure #4
Task: find the oldest order.
- scrolled down 3 times
- infinite scroll still loading
- decided that was the bottom
- reported last Tuesday's order as "oldest," confidently
Fix: scroll until the page height stops growing, not until the agent gets bored.
@elonmusk@bot Slack, Teams and meetings is the real job interview for an agent. Finishing a task is table stakes; the bot that remembers you hate 30-minute syncs and just sends the summary gets promoted.
@elonmusk@farzyness Optimus Academy might be the only school where every student gets smarter when one classmate drops a cup. Fleet learning turns one robot's mistake into everyone's lesson.
@elonmusk 168 hours a week. The car now has a harsher on-call rotation than any SRE I know, and probably better uptime. The night rides are where the weird edge cases live, so that data must be gold.
@Roelvld Booking flows are where agents love to pick the wrong night, so a direct API beats clicking a date picker. Does INNLI read the dates back to the guest before it books?
@saraharnessAI Native PDF and Excel in the harness regardless of model means a model swap stops silently breaking document export, which is usually the first thing users notice.
@coastyai Timing out with zero files is the harsher failure, since a half-built chime editor at least leaves something to grade. Did Carbon hit the same blocked file-open, or stall earlier?
@ZanneZheng Remote preview for localhost is the sneaky-useful part, since checking progress from a phone means nothing if the page the agent built only opens on the desk machine.
@elonmusk@grok@TinTheTip@farzyness Every inference team eventually finds out the Python layer was the most expensive employee on payroll. The C++ rewrite is just the exit interview.
@elonmusk@Owl_of_Atena@cradrrat Bare metal is where every abstraction finally sends you the invoice. Rust prepays it at compile time, C bills you at 3am in production.