Corey James- 246work orders governed, including the programs that built most of this site
- 106hook modules across eleven lifecycle events, each a refusal
- 31,811metered usage entries tied to the work that spent them
- < 5 sfor another worker to resume a failed one from the graph
My long autonomous coding runs all failed the same way. The lead agent had to hold the operating procedure, the plan, every teammate's state and every piece of evidence in its context, and past some horizon it dropped one of them. The work then diverged quietly. I stopped treating that as a prompting problem. It is working-memory overflow, and the cure is to take the memory out of the model. AgentOS is the control plane I built on that diagnosis. The authoritative state of every engineering program lives in a PostgreSQL work graph. Each agent receives a small, freshly rendered, contract-bound job. Policy is enforced by machinery that refuses forbidden actions. Work closes only when an independent reviewer has the evidence. The agents that built AgentOS worked under it.
The idea
A relational work graph is the only source of truth: systems, projects, work orders, acceptance criteria, tasks, evidence, review findings, amendments and cost. Every Markdown view is a generated projection of that graph, and edits to projections are discarded, because two sources of truth give drift a place to hide.
Every agent job is small, rendered at dispatch time from exactly the slice of state it needs, and bound to a schema that says what must come back. Nothing is carried from job to job inside a model's head, so models are swappable and a crash loses nothing that matters.
Rules are enforced by hooks on the agent runtime. A lead agent cannot re-type or edit an artifact; it routes files and renders decisions. A hook may never report an operation that did not happen. Those are refusals, and no prompt has to remember them.
Judgment is independent. Delivered work is checked by quality reviewers, accepted by a gatekeeper that takes only delivered facts, and challenged by an adversarial reviewer on a different vendor's model. The builder never judges its own work.
What I built
A deterministic transition runtime advances work-order state only through compiled transitions, so the same inputs produce the same state and every move leaves a receipt.
A hook layer over the coding agents' own command-line tools turns policy into refusals.
An independent review stack separates builders, quality reviewers, the gatekeeper and the adversarial reviewer into distinct seats, with follow-up review bound to a frozen set of findings so it converges instead of re-litigating forever. The design is in Separation of duties for AI work.
Recovery machinery rebuilds an agent's working picture from the graph after a crash or a context compaction, and replaces a broken session without losing the work order's state.
A managed team runtime launches provider-neutral agent seats in containers through my agent runtime, with signed events and per-seat cost attribution, exposed to agents and operators as MCP tools.
Around those: cost attribution per work order; runtime observability; a helpdesk channel for defects in the harness itself; long-running standing work orders fed by inbound events; and provider adapters that run real interactive sessions of vendors' authenticated command-line tools under my own logins.
Seven questions
An organisation that runs AI workers has to be able to answer, for any piece of work: who did it, why was it allowed, what grounding was used, what changed, what evidence exists, what did it cost, and can you trust it. AgentOS answers all seven from the work graph in seconds, because each answer is a query over rows the system wrote rather than a reconstruction after the fact. A conversation with the platform compiles into two things, a governed execution system and the governance around it: the work order is the source program and the governed execution is the compiled program.
One real program on this platform shows what that graph holds. One work order has been stood down and resumed eight times with the same state producing the same bytes. The economics of that program are in What the work cost, to the penny.
3.197 billion
tokens attributed to the penny
103
distinct status paths taken through the work
Reacting to the world
Every team lead and seat has a mailbox, scoped to its team, with identities derived on the server from the authenticated participant. External systems register as event sources against a work order, so a signed webhook from a payments, email or social-publishing provider becomes a governed event in the owning lead's mailbox and then a finite task. A work order can keep an open-ended envelope, a standing lead that admits tasks as events arrive and brings specialists online on demand, so continuous operation is composed from the same machinery as a bounded program. The idea is Agents that react to the world.
The judgment in it
I made Markdown a projection on purpose, because agents and humans both like editing a plan file and that is exactly where two truths come from.
I forbade the most capable agent in the system from touching deliverables.
It made the lead feel slower and removed a whole class of silent corruption.
I spent weeks stabilising the control plane when a prompt-only approach would have worked for a day at a time. Several of my written rulings from that period, such as anchoring quality checks to lifecycle state and treating prose as a medium for behaviour only, are still the doctrine it runs on.
I held a platform build open for two weeks against the agents' own "done" until every close item had an independent verdict. That one is Closed by evidence.
What it demonstrates
Control-plane and state-machine architecture, PostgreSQL schema design at scale, and the judgment to turn a recurring failure of AI agents into rules a platform enforces. The programs AgentOS has run are the proof: a native port, bit-identical in every parity cell, in under three weeks, an API platform closed all 23 of its close items under independent review, a fine-tuned model with a frozen evaluation.
Stories
Why long agent runs fail · Closed by evidence · Separation of duties for AI work · Directing AI coding agents at program scale · What the work cost, to the penny
Technologies
Rust · PostgreSQL · JSON Schema · Model Context Protocol · Agent runtime hook systems · Multi-agent orchestration · Separation-of-duties review design · systemd