saved
The Company as a Codebase
gist
Frank Chen argues that every workflow should ship two artifacts: today's result and a versioned improvement to the system that produced it. At his company, Foundry sessions close with a writeback pull request into a git-backed company brain, and frontline operators dictate the why behind judgments so an agent can cluster them into doctrine. Autonomy stays inside sandboxes; human gates stay on production merges and shared doctrine, because the result depreciates while the inherited improvement is the company.
ideas
- Two artifacts, one workflow. The result handles today; the writeback changes where the next person or agent begins.
- The writeback is a pull request, not a note. Reviewed, merged, and deployed into ontology descriptions, the prompt builder, and reusable skills. The repo holds doctrine, never raw customer records.
- Build the CI habit before agent volume. Practice first, automation second: by the time agents write into the brain, review gates already exist. Ordinary content self-publishes; humans approve changes to the gate.
- Frontline capture without a keyboard. Operators talk-to-text the why before a job advances; a twice-weekly human gate turns repeated or expensive captures into doctrine the next job inherits.
- Gates that can fail, and boredom. Size autonomy by blast radius. Do not close Carl's diagnosis gate because the agent is usually right. Knowledge systems die from inattention, not disproof.
quotes
“That is the two-artifact rule: every workflow ships the result and a versioned improvement to the system that produced it.”
“The unit of work isn’t the session. It’s the diff the session leaves behind.”
“most knowledge systems die the same death: not disproven, just quietly unattended.”
“The result depreciates. The improvement compounds only when someone inherits it.”