Programs of AI coding agents that finish, prove it, and can be trusted
Rust control planePostgreSQL work graphBuilders and judges from different model families
State outside the model
Work graphthe authoritative state, in PostgreSQL
Jobs · small, fresh, contract-bound
23/23close items, each independently passed
246work orders governed, including every program on this page
45,097state transitions, each one leaving a receipt
1,000+passing trials voided when I found they were testing the wrong thing
AI coding agents can carry substantial engineering only if the output can be trusted. Agents over-claim, drift as their context fills, and optimise for whatever metric is in front of them.
Longer prompts and more human review do not scale.
The method is written down and enforced by machinery I built, and work closes only on evidence.
One long session
rule
Procedure, plan, teammates' state, evidence, decisions: as new work arrives the window slides, and the first thing lost is the acceptance rule. The agent declares done because it no longer remembers what done meant.
One governed job
The work graph holds every work order, task, decision, piece of evidence and cost. Markdown views are projections; edits to them are discarded.
The slice this job needs, rendered at dispatch, with a schema for what must come back. Nothing is carried from job to job in a model's head.
Drift is working-memory overflow, so the cure is less to remember. See Why long agent runs fail.The people who build do not judge
Builders · one model family
B1B2B3
Judges · a different vendor's models
QAGKADVCL
Quality review, a gatekeeper that accepts only delivered facts, an adversarial reviewer, and a closer. Follow-up review is scoped to the findings it was opened for, so it converges.
One swarm phase · 27 gates audited
passed
failed and left failed: workspaces not archived, credential-shaped values in a working document
set aside by explicit decision
Closed as a conditional pass with two failures, rather than narrowing the criteria; the archival tooling was built afterwards.
ActiveKV native portBit-identical in every parity cell, in under three weekszero Python files in the release; everything downstream of a mis-bound component was redone with proof of binding order
ActiveKV performance
24 promotions in four days, each through an intake review; one genuine gain rejected because its cutoff matched the test fixture
API integration platform
held open against the agents' own "done" until the evidence existed
Capabilities modelData, training, a custom decoder, serving and a frozen evaluationunder a single work order with an independent quality seat
Corey James