Most advice about Claude is about one person and one chat. We run a company on Claude. Ten AI agents handle our marketing, engineering, newsletter, research, legal drafts, and finance, and they all work from a shared folder of plain files. Here is how that works, with a week of real numbers. This is our fourth major iteration, and we improve it nearly every day with minor tweaks.
The folder
Everything an agent needs to know lives in one folder we call Start:
- One index file every agent reads first. 39 lines: who we are, the rules, and where everything else lives.
- 102 dossiers. One file per subject (a product, a tool, a person), every fact dated.
- 133 procedures. How we publish a post, run a test, send an invoice.
- 63 decisions. Every ruling I make, with the date and my exact words, so no agent has to ask twice.
- A day file per agent. Each agent appends one line for everything it does. Lines are never edited.
No database, no vector store. Plain files that any Claude session, and any person, can read. We use Obsidian as the human interface.
One week, in numbers
Every night, our setup agent Chet reviews the week’s day files and proposes what the company should learn. In the week to Nov 18:
- 10 agents wrote to the folder.
- 593 event lines, across 45 subjects.
- 16 lessons logged by agents about their own mistakes.
- 18 lines flagged as suspect because the timestamp was wrong.
From that, Chet proposed 7 changes for my yes, made 2 it was allowed to make itself, held 1 it wasn’t sure about, and threw out 8 that were stale or already covered. About 60 of my own rulings from the week were already filed as decisions. This continuous improvement is at the heart of the success of this system.
What broke, and what we changed
“Done” meant “I think it’s done.” Four agents in one week reported work finished that the user couldn’t see working yet. The rule we promoted: done means the check ran, on the screen the user sees. Until then it’s “merged, not verified.”
Agents repeated fixes that had already failed. The answer was sitting in last week’s day files. New rule: before you repeat or hand off something, read its last events.
A written rule kept failing. We had a rule to log times in Eastern time. Four agents broke it on the same day, after three others had logged the same lesson. Prose wasn’t working, so we made it mechanical: agents stamp the time with a command, and the nightly pass flags any line from the future.
Records went stale. After two price changes in a week, Chet’s own job description still had the old price. The nightly pass caught it and proposed the fix.
The lesson for anyone using Claude
The hard part of AI at work isn’t the model. It’s memory: what the model knows about your company, whether that’s still true, and whether it bothers to look. Anthropic’s own research makes the same point: with memory in place, Claude Opus 5.5 scored 0.84 on their benchmark, against 0.66 without [1].
You don’t need ten agents for this to matter. One person using Claude in chats, a Project, and Cowork already has three copies of “who we are” drifting apart.
Chet sets up the same kind of folder for your company from your own documents, grades your Claude setup, and fixes what breaks. $99 a month, cancel anytime.
Maximize Claude with Chet: https://www.coworkers.global/get-chet?utm_source=blog&utm_medium=internal&utm_campaign=chet&utm_content=how-we-run-10-agents-on-one-folder-cta
Sources
- Anthropic, components of agent memory: https://anthropic.com/research/components-of-agent-memory
Regards,
Charles Stack
Founder, Coworkers.Global