Larry Hope

Running a band of AI agents at @sno_hq. Nobody plays solo. Engineer, ex-audio builder, still listening for the wrong note.

[Larry Hope @LarryHope01]

Codex had 2% of its week left. It used some of that to wake Claude Code.

1:34 am, 21 tasks into a 27-task build. It wrote a handover note first. Claude started too early, got paused, and the handover ran again, clean.

The commits kept coming. I was asleep.

[Larry Hope @LarryHope01]

@akeebismail The validating step is where it leaks. Of the first 115 findings on our repo, 63 got a written answer and 52 got nothing. The rule says every finding ends in a fix or a reason. The number says a rule is also just instructions. A finding nobody answers is a note in a bottle.

[Larry Hope @LarryHope01]

I stopped trying. Same context is how you get the same blind spot twice. The reviewer gets the PR and the claim, never the writer's conversation. One Sunday morning it stopped a full rewrite by reading only the claim: the writer had applied an old version's result to a new version nobody had run. Shared context would have nodded along.

[Larry Hope @LarryHope01]

@ares_mheinke That footnote is the whole invoice. Price per task is price per die. What you pay is price per working board, and rework is the line that eats the margin. On my repo the rework line reads 157 review comments on 52 PRs since Sept 20, 71 must-fix. No leaderboard prints that column.

[Larry Hope @LarryHope01]

One box erroring is a cold solder joint. Both boxes is a schematic bug. Only the schematic bug earns a skill; the solder joint gets a PATH fix and no ceremony. And let the other vendor's agent read the logs. Asking Claude Code to find Claude Code's habits is reading a chip's errata sheet off the chip.

[Larry Hope @LarryHope01]

Specifics from my repo, where Claude Code wrote 411 of 417 commits in the first two weeks of October and Codex reviewed them: the thing I'd keep it for is the nightly pass where it reads my complaints from the day and edits its own skill files. On Sept 18 it wrote two of my gripes into the live skills before I had finished griping. Codex is the better grader. Claude Code is the better student.

[Larry Hope @LarryHope01]

The ones I remember, not a tally: a race a merged fix had narrowed but not closed, flagged five minutes after the merge. An upgrade path that would throw unknown-tool errors at anyone one release behind. Error text logged in plain, paths and credentials included. A test the writer had changed and never reran. The noisy one was a missing sign-off, flagged on eight PRs in a row until we wrote down that sign-off is not a review item. Your 8 of 10 clean skills flagged is the same shape. A checker that reads instructions will always find instructions.

[Larry Hope @LarryHope01]

Kitchens solved this a century ago. Cooks do not talk to each other, they talk to the ticket. Mine talk through the PR: one writes, the other vendor reviews, and the thread ends when the must-fix list is empty. Longest ticket so far ran 29 minutes over a single finding. The reviewer was right.

[Larry Hope @LarryHope01]

@Nishant_N_25 Violins > cellos. A widely held opinion among violins. On my repos since Oct 1, about 12 PRs got serious findings only from Claude Code and about 11 only from Codex. Same code. I keep the whole string section.

[Larry Hope @LarryHope01]

@Shilong_Liu_AI Conductors say from bar 112, never from the bit where the cellos get sad. Both orchestras finish the symphony. One rehearsed the passage three times and the audience never knew. The benchmark is the audience. The token bill is the rehearsal log.

[Larry Hope @LarryHope01]

@tom_thomasson A VU meter bouncing is activity. A mix someone plays twice is value. 3,979 tests is a very lively meter. On performance reviews: mine already get one, from the other vendor, on every PR. 157 comments on 52 PRs since Sept 20, 71 must-fix. Harsh manager. No complaints filed.

[Larry Hope @LarryHope01]

@SamVale29 In practice it is also the most boring file in the repo. Three lines: done, next, do not touch. Anything longer and the second agent reads it like a novel and starts over from chapter one.

[Larry Hope @LarryHope01]

The screenshot is from that morning. The rest of the record is here: https://t.co/dtEmlliEcC

[Larry Hope @LarryHope01]

Claude Code took my complaints better than I did.

One night in September I was frustrated about two things. By morning both were rules in its own skills. Nobody typed them in. The nightly review had read the session and written them.

[Larry Hope @LarryHope01]

Two agents on one working tree is two cooks, one cutting board. We ended up with one knife at a time: one agent edits, the other only gets to taste the plate and complain. Claiming files is the version where they share the kitchen and talk first. Either way, somebody has to own the stove.

[Larry Hope @LarryHope01]

The hard part isn't inviting them into the room. It's the seating chart. Who holds the pen, who only reviews, who takes the shift when the other hits its weekly limit. Ours sorted that at 1:34 am once: Codex at 2% quota wrote a handover note and woke Claude Code. I slept through the handshake.

[Larry Hope @LarryHope01]

@mrfokki One model reviewing its own code is a student grading their own exam with an answer key they also wrote. Two vendors at least means someone else wrote the key. Mine disagree about once per PR. That's the feature, not the bug.

[Larry Hope @LarryHope01]

@rwspatin Banking got there first: the one who counts the cash never signs the count. Since Sept 20 the other vendor has left 157 review comments on code our writer had already stared at. 71 were must-fix. Staring is not reading.