1,006 commits in 100 days. about 90% were written by agents

Pixelwand crm, 70 working days, and a job that changed from writing code to reviewing it. What the commit log shows, and what it hides.

1,006 commits. 70 days with at least one commit. june 19 to october 1, one crm.

about 90% of that code was written by agents. most of my time went into reviewing it.

and i think the number is small. i want around 3,000 commits in three months.

the crm

pixelwand crm is a light weight crm for fast paced teams. whatsapp, email, sms and calls, all inside one tool.

what the log says

i sorted the commits by prefix.

491 fix
252 feat
94 refactor
52 chore

half of it is fixes. a crm is phone numbers, imports, whatsapp templates and permissions, and every one of them breaks somewhere the happy path never goes.

what the log hides

37 commits have a Co-Authored-By: Claude line. that is not the real number.

i mostly use codex, and codex does not add a co-author line. claude does. it likes to flex. so a git log of this repo makes it look like 4% agent work when it is closer to 90%.

if you want to know how much of a repo an agent wrote, dont trust the trailers.

the two days that look like my best

aug 26 and 27: 273 commits.

almost every one looks like this:

fix(types): resolve type errors in CallLogModal.vue
fix(types): resolve type errors in TaskPanel.vue
fix(types): add PendingInvite type and parameter types in Invitations

on those days we moved the untyped vue components to full type safety, to make the code easier to maintain and to catch errors before they reach production.

before: about 3,500 typescript errors. after: zero.

by commit count, those two days were 27% of the whole project. by the work they were one cleanup job, done one file at a time.

where agents fall short

not on rules. the agents follow the rules file.

the problem is size. give an agent a big task and it will not do all of it. it skips parts. most big tasks took me two to four passes, repeating the instructions each time.

so the job became: ask, review, ask again for what was skipped, review again.

the review

nothing merges until the tests, typecheck, format and lint all pass. then a human reads it.

that human is me, and that is where the time goes.

the rules file

AGENTS.md is 204 lines. five rules sit at the top and override convenience:

frontend and backend are coupled only by a typed contract
thin api, deep service
errors should be loud
reads are pure
no n+1, no work on import

it also lists the oversized files and says not to extend them in place. extract the slice you touch.

how the power dialer got fixed

the power dialer started as vibe code. we built it fast for one customer.

then we engineered it. it is now an api file of 542 lines on top of a package of six service modules: sessions, queues, leads, dispositions, analytics and common code.

billing.py is the next one. it is 4,139 lines today.

what i would change

nothing.

if i started again in june i would do the same: ship the fast version for the customer who needs it, then pay it down properly. the order matters. the cleanup is easier when you know what the thing actually does.

the number

1,006 is not impressive to me. what i trust is whether the next week is faster than the last one.

the log says the bottleneck is no longer typing. it is review. so next, how do i get an agent to finish the whole task in one pass?