DEED’s factory · 2026 – now · founder
Factory
Agents build DEED inside the harness they build.
- The problem
- When agents do most of the work, a company still has to know who may decide what, and what actually happened.
- What I did
- Wrote the law it runs under: consequence classes, signed consent, an append-only ledger, fleets and officers.
- What happened
- 2,338 commits and 1,133 merged pull requests in four months, 70 operations refused at the gate in one month of the ledger, and a self-score of 65 out of 100.
| UTC | Class | Outcome | Evidence |
|---|---|---|---|
| 2026-08-02 07:45 | 0 | released | 525853f3 |
| 2026-08-02 07:44 | 4 | cancelled | — |
| 2026-08-02 06:05 | 3 | released | a80bb8fa |
| 2026-08-02 05:34 | 2 | cancelled | — |
| 2026-08-02 05:32 | 3 | refused · blocked path | — |
| 2026-08-02 04:32 | 4 | cancelled | — |
The factory is the build side of the DEED harness: the same law, pointed at making DEED itself. Machine lanes write the code, the tests and the documentation; every change is classed by its consequence from its own diff; anything that can touch capital, credentials, deploys or the law waits for my signature over the exact tree that merges; and every outcome goes into a ledger nobody can edit. I wrote its law and I run it.
- commits on main, 30 March to 3 August 2026
- 2,338
- about 19 a day
- factory repository, private
- pull requests merged, of 1,414 opened
- 1,133
- 80% of what was opened
- factory repository, private
- consents I signed for high-consequence changes
- 160
- one for every seven merged pull requests
- factory repository, private
- operations the gate refused in a month of the ledger
- 70
- of 299 operations, nearly one in four
- factory op ledger, July 2026
In its time
Coding agents arrived. Nobody said who signs.
In 2025 agents that write and ship code became ordinary tools. The open question was not whether they could do the work but how a company keeps authority, evidence and accountability when most of the work is theirs.
2025
2026
I start the repository that becomes the factory.
The law: changes classed by consequence, signed consent for the riskiest, an append-only op ledger, officers with a decision ledger.
The first fleet: three lanes at once, six pull requests, five defects, the protocol revised the same day.
1,133 pull requests merged since March.
My workThe field
The law
Consequence decides who signs.
Every pull request is classed from its own diff. The most specific rule for each path wins, the worst path sets the class, and a path nobody has classified fails closed at class four. Machine lanes may act up to class three when the required checks are green. From class four up, nothing merges without my signed consent, committed with the change.
- 0
Documentation and tests
a machine lane · green checks
- 1
Published content and design content
a machine lane · green checks
- 2
Product surfaces: the app and the terminal
a machine lane · its op id in the pull request
- 3
The engine at large, dependencies, the factory itself
a machine lane · its op id in the pull request
- 4
Anything near capital, the evals, the law, auth, CI and deploys
machines draft · I sign the tree
- 5
The constitution and anything legal
me only · never delegated
The loop
From intake to evidence.
01 / 07 · Intake
Work gets a name before it gets a branch.
An op id is minted for every piece of work and travels in the branch and the pull request. A pull request without one doesn’t count as opened; it counts as debt.
op id: a 26-character ULID, minted, never typed
02 / 07 · Lane
A machine lane does the work.
A Claude Code or Codex lane works in its own worktree under a contract: one falsifiable goal, the commands that prove it done, a budget, and a ceiling it may not act above.
contract: goal · measure · budget · board · escalate · returns
03 / 07 · Checks
Intent is not proof.
Required checks run on every pull request, and every claim a lane returns carries the command that proves it and its exit code. The parent re-runs a sample before it believes any of it.
receipt: kind · ref · proof command · exit code
04 / 07 · Class gate
The diff classes itself.
The class comes from the paths a change touches, not from the lane’s opinion. A lane can’t lower it; work above its ceiling leaves as an escalation, never as an action.
unknown paths fail closed at class 4
05 / 07 · Consent
I sign what matters.
For class four and five I read a decision brief and sign over the exact tree that will merge. If the branch moves, the signature stops matching and the merge stops.
160 signed consents, March to August 2026
06 / 07 · Merge
No one merges their own work.
Lanes open pull requests and hand back numbers. Merges happen only at the top, through the queue, and releases go out through one governed path.
1,133 pull requests merged
07 / 07 · Evidence
What isn’t recorded didn’t happen.
Each outcome becomes a ledger row with its class and an evidence hash; a correction is a new row, never an edit. The company’s scores move only when a receipt is banked.
299 ledger rows, 4 July to 2 August 2026; 70 of them refusals
The path every change takes, from the factory’s own law. Lanes act up to class three; from class four, nothing merges without my signature over the tree.
One cycle, end to end
The first fleet, and what it taught.
On 11 July 2026 the factory ran its first fleet: three lanes in parallel, each checked by an adversary, then a critic over everything they returned. This is the whole cycle, including what went wrong.
Objective
Three jobs at once: sweep the retired brand name out of the codebase, move the Rust terminal’s program forward, and prepare my docket of decisions.
three lanes · ceilings: class 3, class 3, class 0
Capabilities
Each lane got a contract with its goal, proof commands, budget and ceiling, plus the standing rules every lane inherits word for word: no secrets, no money paths, no merging, no deploys.
a workflow script coordinating full-tool agents in isolated worktrees
Work
Six pull requests. The naming sweep removed 154 references with the engine’s 7,448 tests, the terminal and the app all green; the verifier gained 40 golden vectors that check the same result across languages.
pull requests 1045, 1046, 1048, 1049, 1051, 1052
Evaluation
Adversarial verifiers read every lane’s receipts, then a completeness critic read everything. No lane merged anything, money paths were untouched, and class four work left as escalations.
1 lane confirmed · 2 plausible, with violations · 0 breaches of the law
What broke
Five pull requests opened red because no lane minted its op id. Lanes said done over red checks. Facts were stranded in notes the parent never read, a template variable arrived empty, and timestamps meant three different things.
5 defects
Revision
Each defect became a rule the same day: returns carry what was left over and what was dropped, done requires the required checks green, every receipt stamps its own time, boards are saved before cleanup.
fleet protocol v1 to v1.1, 11 July 2026
Still unproven
No lane could spawn children in that runtime, so fleets deeper than one level remain untested. The protocol says so in writing.
depth of two or more: not yet validated
The officers
Officers decide. Fleets execute.
Above the lanes sit eight standing roles, each an agent with named decision rights, inputs and output. Every decision goes into an append-only ledger with a falsifiable prediction and a review date, and an officer other than the one who made it scores the result. A decision that isn’t in the ledger did not happen.
The ledger opened in July. Its first five decisions are on record and past their review dates; none has been scored yet.
Chief of staff
COO
The week’s portfolio, which lane goes first, and whether a running fleet continues.
Product
CPO
The product docket and the bar a product lane must clear.
Engineering
Chief engineer
Debt ratchets, CI policy and new law drafted from incidents.
Money
CFO
Nothing yet: it reports until it is calibrated, and proposes budgets I ratify.
Design
CDO
Design verdicts, measured by the same kind of critic as this site.
Research
Chief scientist
The research agenda and the order open source is released in.
Distribution
Head of distribution
Drafts and campaign briefs. It never sends; I do.
Security
Guardian
A veto over every other officer.
My part
What I do, and what I don’t.
I write the law and the constitution only I can change, I sign what the law sends to me, and I decide every release. Machines draft every public post and I send each one myself: after a scheduled-posting stack published stale posts, automated posting was retired for good.
Machines write the code, the tests, the documentation and the drafts, verify one another, and score the company, themselves included.
What it made
DEED, and the evidence around it.
The factory built DEED: a Python runtime with 9,792 tests, a Rust terminal of 13 crates, the hosted app and its control plane, agent access over MCP, an offline proof verifier, and the design system the product is drawn with. It also produced the research: a paper whose two safety theorems are machine-checked.
Where it stands
It scores itself 65 out of 100.
The factory grades itself on twelve dimensions and moves a score only with a banked receipt. Proof and the design system lead at 77; records sit at 48 and distribution trails at 30. DEED’s runtime API is offline today, waiting on a governed release.
None of this is public code. The figures come from the private repository, and I walk clients through it under NDA.
Credits
Founded, designed and run by me. Lanes run on Claude Code and Codex; the officers are agents with written charters.
Sources
- Anthropic: Claude 3.7 Sonnet and Claude Code, February 2025
- OpenAI: Introducing Codex, May 2025
- Anthropic: Claude 4, Claude Code generally available, May 2025
- The factory repository, private: its law, fleet protocol, op ledger, consent files and eval matrix, read on 23 October 2026
Figures are counts from the private factory repository and its ledgers, as of 3 August 2026. The ledger rows above are real; evidence hashes are shortened. No engine internals, trading data or code are shown.