The agent fabric, with the lid off.
There is no control plane. ARCHON plans the DAG, dispatches four isolated workers, and checkpoints every step — entirely on your cluster. Six kernel modules, six published contracts, nothing sealed.
brew install archon-runtime/tap/archonEvery run leaves a receipt.
Submit a workload and watch the kernel ingest it, compile a DAG, dispatch four isolated workers, and commit a durable checkpoint. The event bus is not decorative — it is the same stream archon trace reads.
Ingest
Normalize the workload, content-address every page, and open a run ledger.
Plan DAG
Compile the goal into a typed, acyclic plan. Pin a model and a budget per node.
Dispatch workers
Place ready nodes on four isolated workers. Each lease is scoped to one worker.
Commit checkpoint
Fsync state, output hash, and idempotency lease. The resume point is now pinned.
Six frames hold
one row.
The six vertical threads are the kernel modules. Each red pass is a worker laying down one row — and no row is woven until every frame has held it. The cloth is the committed run; a stopped run resumes at the last locked row.
Six modules. Six published contracts.
A module is not a folder. It is an input, an output, a stated semantic, and a signature you can pin. Select a node to read its contract. All six ship in the same 38 MB binary.
A runtime you can audit.
No hosted tier, no closed agent harness, no hidden scheduler. These are the parts, their defaults, and how to verify each one before you trust it with a run.
| Subsystem | Implementation & default | Verification |
|---|---|---|
| 01Runtime core | Rust 1.82, single static binary, 38 MB. No runtime dependencies, no daemon required. | verify sha256 + cosign on every release |
| 02Scheduler | DAG-aware, work-stealing, fair-share per tenant. Backpressure engages at 92% queue depth; no priority inversion between tenants. | verify archon sched --explain |
| 03Isolation | gVisor by default; Firecracker optional. One worker per lease, capability-scoped tool calls, no ambient credentials. | verify archon worker --inspect |
| 04Durability | Per-step checkpoint with WAL and exactly-once resume. Stores: SQLite (default), Postgres, NATS JetStream. | verify kill -9 test in CI, nightly |
| 05Transport | QUIC peer mesh with mTLS. No control plane; membership and ownership leases only. | verify archon peers; tcpdump shows no vendor host |
| 06Kernel modules | Six contracted modules, semver'd, ABI-stable within a minor. Contracts are generated from the source, not written by hand. | verify archon contract diff v1.4.1..v1.4.2 |
| 07Model router | 14 providers, policy-based failover, token accounting written per node so a resumed run is billed once. | verify archon route --dry-run |
| 08Governance | Apache-2.0, open governance, no CLA. Every design decision lands as a public RFC. | verify github.com/archon-runtime/rfcs |
Three commands to a running fabric.
Install one binary, declare the fabric in archon.yaml, then boot it on your own node. Select a step to preview its exact input.
Deployment, answered plainly.
The questions an SRE asks before the second meeting. If one is missing, the ops guide is public and the issue tracker is the front door.
No. Archon has no vendor control plane and no telemetry endpoint. The only egress is the model providers and tool hosts you name in archon.yaml. Point the router at local weights and archon up runs with the NIC unplugged.
Every step commits a checkpoint — state, output hash, and idempotency lease — to the configured store before the next step may start. If a worker dies, the run resumes at that committed step. Side effects are keyed by lease id, so a retry cannot double-charge. The CI suite kills workers at every step boundary.
A workflow engine replays history you define up front. Archon's Planner compiles a typed DAG from a goal, the Model Router pins a model per node, and the Scheduler places ready nodes on isolated workers. Durability is the floor, not the product.
Yes. Mirror the release into your registry, set registry.mirror, and route models to vLLM, Ollama, or any OpenAI-compatible endpoint on your network. No license check, no phone-home, no clock-based expiry.
One static binary, 2 vCPU and 2 GB RAM. SQLite is the default checkpoint store; Postgres or NATS JetStream for HA. The three-node reference cluster runs in 6 vCPU total.
Peers form a QUIC mesh. Each run has an ownership lease with a monotonic epoch; a node can only advance a run while its lease is current. If the owner disappears, the lease expires and another node resumes from the last committed checkpoint. Raft is used only for cluster membership — never to schedule runs.
Windows ships a native amd64 build and runs under WSL2. gVisor is Linux-only; on Windows, workers fall back to process isolation with Job Objects and a restricted token. The parity matrix is published with every release.
Own the machine
that runs your agents.
One binary. Six published contracts. Zero control planes. Install it on the node you already trust.
curl -fsSL https://get.archon.run | sh