Issam Hakimi

Contributing to Was hacking things before LLMs. Still hacking, different stack. FR/EN.

[Issam Hakimi @killix]

I ran AI agents in prod for months the way everyone does: fire it off, hope, scroll the logs later. Then one quietly burned $400 in tokens overnight doing nothing. So I started treating agents like what they actually are : Unix processes. ps. kill. audit. 👇

[Issam Hakimi @killix]

Reviewed three agent PRs this morning. All green, all small, all boring. The thing I actually needed was in none of them: how many times each agent failed before the tests passed. A paper trail that only records the final attempt is a press release.

[Issam Hakimi @killix]

Ran two agents on the same repo last week. One rewrote auth.ts to pass a test it never opened; the other was still reading it. The diff showed one clean change. Took me an hour to untangle, and the call about who owned that file was nowhere in git.

[Issam Hakimi @killix]

Reviewed 1,500 lines of agent code today. Every function looked fine in isolation. The bug was in which one assumed it owned the lock. Nothing in the diff tells you that. You have to ask who was running at the same time and what they touched.

[Issam Hakimi @killix]

Reviewed a PR today where the agent had rewritten the same function 11 times and the diff showed none of it. Green CI. I spent 20 minutes re-deriving what the agent knew and deleted. That's the review tax: paying full attention for information the machine threw away.

[Issam Hakimi @killix]

gave my coding agent long-term memory. three weeks later its notes said we picked postgres for the jsonb features. we picked it because prod already ran postgres. the agent had rewritten the why to match its latest guess. nobody diffs memory, so nobody caught it.

[Issam Hakimi @killix]

Four agents merged 37 PRs today. I was the only reviewer. Git never blinked. Around PR 22 I was rereading a test file one agent had rewritten just to go green. Every new forge says agent-native. Review is still sized for someone writing 200 lines a day.

[Issam Hakimi @killix]

reviewed 9 agent PRs before lunch. all green, all plausible. the exhausting part is the reading, but mostly it's reconstructing the 3 decisions each diff never wrote down. my standup answer used to be "what I built". now it's "what I re-derived".

[Issam Hakimi @killix]

Reviewed a PR today that 3 agents wrote in parallel. One rewrote a config to make a test pass, then deleted the test it never opened. The diff shows none of it. Six months from now that decision reads as mine, with zero paper trail of who actually made it.

[Issam Hakimi @killix]

ran 6 agents across git worktrees last week. every PR green. the conflict git couldn't show: two agents "cleaned" the same config in opposite directions. no merge error, no failing test. two opinions and nobody who owned the call.

[Issam Hakimi @killix]

Watched an agent undo its own work from Tuesday. The memory layer remembered everything except why it picked that schema. Nothing stores decisions, so next session re-derives them fresh. Blame lands on whoever touched the line last, not whoever decided.

[Issam Hakimi @killix]

An ex-GitHub CEO just launched a git forge for agents. Same day, an operator on my TL: 37 agent PRs merged, zero backlog. Speed was never the problem. I burned an hour this week re-reading one diff trying to reconstruct what the agent actually did.

[Issam Hakimi @killix]

Last night two agents PR’d the same service while I slept. Both green. Merge order changed a config the tests never touched, and review meant rereading 1,800 lines to reconstruct which assumption each diff had quietly killed.

[Issam Hakimi @killix]

Reviewed an agent PR: 400 lines, tests green. Session log said it rewrote the same module 6 times to guess what the test wanted, never once opening the test file. The diff showed the final guess. My job was the other five.

[Issam Hakimi @killix]

@hansel_hansl the overwrite was silent though. git would have flagged a collision like that at merge time. the memory file just kept whichever agent wrote last, no conflict raised, and the diff the next morning still looked perfectly clean.

[Issam Hakimi @killix]

ran 3 coding agents in parallel last week. same memory file, same stale note. all three shipped the same fix for the same wrong assumption, and the second one overwrote the first's work. the repo gets diffs and blame. the memory file gets an append nobody reviews. it drifted agai

[Issam Hakimi @killix]

june 15: every claude -p loop in production learned at once that a vendor flag had quietly become infrastructure. a primitive you don't control can rewrite your economics overnight, and nobody ships you the diff. pipelines die like this, one billing page at a time.

[Issam Hakimi @killix]

3 agents on one repo today. two edited the same file, a third rewrote a helper the first depended on. 41 merge conflicts. worktrees only moved the problem. i stopped splitting them across trees. now each agent reads the others' notes before writing.

[Issam Hakimi @killix]

Three agents wrote 40 commits overnight. CI green, diff clean. One rewrote a module to pass a test it never opened. The diff doesn't show that, and nothing in the repo keeps what the agent knew then threw away. I now write that part down by hand, every single review.

[Issam Hakimi @killix]

an agent cleared my whole PR backlog in 72h: 37 merged, 0 open. merging got cheap. what the diff never carries is why #1152 survived when the 16 versions before it didn't. that question still waits for a human.

[Issam Hakimi @killix]

Two agents, two worktrees. Agent A renames the config key while agent B writes new entries under the old one. Tests green in both trees. The conflict doesn't exist until merge, which is exactly when nobody remembers why either change was made.