Kill a process the moment it stops shipping code

On Clarity, I built and ran a multi agent workflow: four parallel sessions covering planning, UX, architecture and build, each in its own git worktree on its own branch, sharing one repository, with role derived from the directory name and enforced by a hook that refused writes to documents another session owned.

The measurement

In its last fifteen merges, the workflow produced zero lines of source, and five of seventeen pull requests carried no work at all, only messages between sessions. There was one developer on the project. I dismantled the workflow in a day, and deleted 2,833 lines of planning documents down to one 215 line context file, keeping only what was an open bug or a fact that had cost real money to learn.

Why this matters more than building the system

Process cost is a product cost. Most builders cannot show the judgement to shut down their own process the day it stops producing software, rather than the day it fails outright.

What replaced it

Three rules, held as current best guesses rather than conclusions:

• An authority order. The repository or the database wins over the project document, and the project document wins over the summary. A disagreement is resolved by querying the system, not by reasoning about which document sounds more recent.
• Cite, never restate. If a fact lives in one document, other documents reference it by section instead of reproducing it. Verified by grepping for distinctive figures across documents and finding zero duplicate hits.
• A written record of every contradiction, and how it was resolved, so a later drift happens against a record instead of quietly re-resolving itself into a new wrong answer.

Underneath all three: documentation for a human reader and context for an agent are two different artifacts with two different budgets. Writing one and expecting it to do both jobs is what produced the planning-document overgrowth in the first place.

Adversarial review scoped by ownership is the same independence principle behind independent verification: a verifier should never be spawned and graded by the party it is checking.

See Clarity for the source project and the published post telling this story.