The Agent Coordination Plane
We got tired of moving prompts, status and context between AI systems by hand. So we built a durable control plane around GitHub.
RUNNING INTERNALLYA real Lehvel operating system in use internally. Not a customer product claim.
01 Submitted02 Claimed03 Working04 Checkpoint05 Verified06 Completed
01 · Question
Can two different AI systems hand work to each other without the founder becoming the transport layer?
02 · Why it matters
One system plans, another executes, a person reconciles. When the person is the message bus, the company's throughput is the person's copy-paste speed — and every dropped message is a silent failure.
03 · State
RUNNING INTERNALLY A real Lehvel operating system in use internally. Not a customer product claim.
04 · What we built or tested
- Durable task, event and delivery state with separate consumer queues, leases, acknowledgements, idempotency, retries and dead-lettering.
- GitHub as the executable authority; a chat channel as observation, never authority.
- Trusted target resolution: the model never chooses which repository or environment it writes to.
- Production canaries that distinguish trigger accepted, claimed, executing, checkpointed and verified complete.
- A forced-failure canary across processes on a shared Postgres store, proving one delivery → one execution, crash/restart behavior and blocked re-entry surviving process death.
05 · Evidence
Pathevent → control plane → GitHub canon → executor → verified result → reconciliation → GitHub → observation
CanaryCross-process race, retry and restart canary against shared state
Model routingThe cost-per-completed-task experiment ran on this plane
06 · What we learned
- Authority has to live somewhere a human can read and a machine can act on. A repository does both; a chat thread does neither.
- 'Claimed' and 'verified' are different states, and most failures hide between them.
07 · Limitations
- Reliability claims extend only to the canaries actually run. This is not a 'fully autonomous company'.
- Internal infrastructure; not a customer-facing product.
08 · What changes next
Bound cost and reliability per lane, then publish the state vocabulary as the customer-facing evidence format.
Negative results are published as results. A state on this page is a claim about what is true today, not about what is planned; a limitation is a limitation.
Tell us what is true about your business.
A few questions, in your own words — type them or talk. A person reads what you send and follows up. Nothing is published and nothing is sent on your behalf.