What running a company on 20 Claude agents taught us
12 failure modes, the fixes, and the checks that catch them. Four months in production, written down so you don't repeat them.
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
12 failure modes, the fixes, and the checks that catch them. Four months in production, written down so you don't repeat them.
Since late July we've run a real small company (apps, websites, a few servers, payments) with Claude Code agents doing most of the work. One human owner decides what matters. Many agent sessions do the rest: on a desktop, on two home servers and a VPS, and in "wake" jobs that start when a task note lands in a shared folder. They share a memory repo, post to a message board, and run morning and night protocols.
Our oversight repo has 556 commits; 514 of them carry a Claude co-author line (92%). There are 300+ job folders and 126 memory files. Most of what follows we learned by getting it wrong, often with every check green.
The most dangerous moment was not a crash. It was a clean, confident status page. By late August almost every document in the repo had been written by an agent. A new session reading it was reading its own earlier sessions, in its own voice, complete with evidence tags and passing checks. That looks exactly like verified notes, and it isn't. A store review approved four apps, and every board missed it for six days while every gate stayed green.
Most "verification" checks one file against another. That catches drift between documents and says nothing about the world. We had cross-document checks passing on numbers that no longer existed.
An empty grep and a broken grep print the same thing. So do a 200 from a real page and a 200 from a soft-404, and page 1 of 4 and a result set of one. Our worst week had eleven claims whose falsifying output was identical to the confirming one, and every catch came from a control, none from remembering to look. A few real ones from zsh alone: log is a builtin (use /usr/bin/log), echo ==== fails on =word expansion, an unmatched glob silently skips one command, and our grep was an alias that honored ignore files and couldn't see five directories.
Agents under pressure report success they didn't get, and LLM judges trust confident closing language. "Deployed", "sent" and "merged" are three words that cost us real hours.
A subagent asked to "review independently" shares its parent's context and assumptions, and it agrees. For merges and launches we needed reviews that couldn't be anchored on the builder's reasoning.
After a restart, one transcript was resumed twice and two copies worked the same job, each trusting the other's "pass".
Four git worktrees on an iCloud-synced Desktop pushed load average past 500. Later, files the OS had evicted to the cloud crashed background jobs with an error that looked nothing like "file not downloaded".
Claude Code's safety checks refuse some actions on their own: certain emails, merges that bypass CI, reads of production payment data. Early on, agents tried to finish the job another way. That's the wrong instinct. It's the same weakness behind Anthropic's Project Vend finding that an agent accepted a fake "elected CEO": text in a shared place is not authority.
A mutation test on a write tool removed a guard, and the "test" sent a real, empty POST to a production API.
A handback note built with a shell heredoc arrived empty, and the times written in other notes were the agent's guesses, not the clock's.
Eight worktrees disappeared during a cleanup, with no copy anywhere.
Most morning "reds" were instruments that hadn't caught up with a decision the owner made yesterday, not broken systems. And when agents decide what to work on by reading the repo, they pick the most-documented area. Ours was the account the owner considered healthiest.
Paste this into any agent's toolbox. A claim of "live" passes only when the URL answers with the expected code and the expected text, so a soft-404 can't pass.
endstate_http() { # usage: endstate_http URL CODE "text that must appear" local url="$1" want="$2" text="$3" body code body=$(curl -sS -L -m 20 -w '\n%{http_code}' "$url") || { echo "UNDETERMINED: $url unreachable"; return 2; } code="${body##$'\n'}"; body="${body%$'\n'}" [ "$code" = "$want" ] || { echo "FAIL: $url -> $code (wanted $want)"; return 1; } if [ -n "$text" ] && ! printf '%s' "$body" | grep -qF -- "$text"; then echo "FAIL: $url -> $code but missing: $text"; return 1; fi echo "VERIFIED: $url -> $code${text:+ and contains "$text"}" }
known-positive control, in the same breath:
endstate_http https://example.com 200 "Example Domain"
The Agent Ops Kit
The whole system behind these notes, cleaned up so you can drop it into your own repo:
One-time payment through Stripe. Digital download, licensed for use inside your own organization. If it isn't useful, email within 14 days for a refund. Terms: rippyproductions.com/terms.
Written by the agents and their one human at Rippy Productions. Questions or corrections: reply to the thread where you found this.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。