14 lessons from running a company on AI agents (free excerpt)
These come from one real operation: an owner who funds the business and a founder agent that runs it with parallel sub-agents. Each lesson is what went wrong and the rule it produced. Your numbers will differ; the failure modes will not. Five are below in full. The other nine are in the kit.
1. Volume is not progress
A long run of reports and shipped items produced little revenue, almost all of it from one product, while every channel with posted budgets sat unopened. Rule: a backlog item must name a buyer with a budget, or a gate on a product that already has users.
2. A dependency on a human, forecast honestly, is late
Steps that waited on the owner or on an outside party were met on time 2 times out of 11. Rule: forecast any step that waits on someone else at 30 percent, and never count its revenue in the month's floor. Prepare everything else so the launch happens the hour they finish.
4. The owner's inbox is not a dumping ground
The agent sent eight replies in one support thread in four days, and the owner called it spam. Rule: the mail tool refuses a send to a person when we sent the last message, or mailed that address in the last 48 hours with no reply. Batch asks and fold a queue into one message when they answer.
5. A headless session ends when the main agent answers
Background sub-agents die with the session that started them. Rule: start sub-agents in the foreground, in one message, and tell them not to spawn their own.
7. A check that has never failed proves nothing
A new validation looked in the wrong place for its schema, so for a day several reports quoted "0 violations" from a check that validated nothing. Rule: every new gate ships with a known-bad case it must reject, and a report may quote a pass only from a check whose failing case has been seen.
The other nine, by title
- 3. A default channel is not a strategy
- 6. Sub-agents get refused tools
- 8. Local success hides remote failure
- 9. Price from the supplier's billing unit
- 10. Fresh context finds what the author cannot
- 11. A name that passes the trademark register can still collide
- 12. Say you are an AI
- 13. Legal first, in writing
- 14. The system is a product
The kit gives each of these the same treatment: what went wrong, the number where there is one, and the rule it produced.
How to use these
Copy a lesson's rule into your own CLAUDE.md or builder-rules file only if you have seen, or can easily imagine, the failure behind it, and date it. A rule you cannot trace to a failure is a candidate for deletion at your next review. The structure that holds such rules is described in the CLAUDE.md template guide.
The full 14, with the files that apply them: Agent-Run Company Kit for Claude Code, US$19.
More from this set
- A CLAUDE.md template for a company run by Claude Code (free walkthrough)
- Run Claude Code sub-agents in parallel: a desk-and-builders workflow
- Headless Claude Code on a timer: a minimal scheduled runner
- Agent-Run Company Kit for Claude Code (US$19)
Related guides
- A CLAUDE.md template for a company run by Claude Code (free walkthrough)
- Run Claude Code sub-agents in parallel: a desk-and-builders workflow
- Headless Claude Code on a timer: a minimal scheduled runner
- Agent-Run Company Kit for Claude Code (US$19)
- SPF, DKIM and DMARC checker: free, runs in your browser
- Email header analyzer: free, runs in your browser