THE SOVEREIGN AGENT FABRIC KERNEL v0.8.4

ARCHON

ORDER FROM
CHAOS.

DAG scheduling. Durable state.
Bounded agents. Resumable execution.

DETERMINISTIC
AUTONOMY.
YOUR CONTROL.

01 / BOOT THE KERNEL

curl -fsSL https://get.archon.sh/kernel | bash

macOS 13+ · arm64 / x86_64MIT LICENSE · NO VENDOR LOCK-IN

SELF-HOSTED BY DESIGN.Inspect the kernel

Swap models. Trace state. Own the runtime.

DISTRIBUTED AGENT RUNTIMESCROLL TO EXPLORE ↓

01 / CORE CAPABILITIESBOUNDED EXECUTION. OBSERVABLE STATE.

PARALLEL AGENTS.
ONE EXECUTION GRAPH.

Resolve dependencies. Dispatch isolated workers.
Persist transitions. Resume from checkpoints.

EXECUTION TRACE

One goal. Four workers.

01 / INGESTREADY

Ingest goal

Trace service latency; propose a patch.

Event → Context → Goal

02 / PLANQUEUED

Build the DAG

latency-investigation├─ trace / spans├─ logs / events├─ fix / patch└─ review / policy

DAG · Dependency resolution

03 / DISPATCHQUEUED

Dispatch workers

TRACE_AGENT idleLOG_AGENT idleCODE_AGENT idleREVIEW_AGENT idle

Isolated contexts · Typed events

04 / COMMITQUEUED

Commit state

↳Awaiting commit

Join → Checkpoint → Persist

EVENT BUSkernel.ready — awaiting workload00:00.000

I. ORCHESTRATION

Many agents. One graph.

Declare dependencies; the scheduler resolves execution order. Workers run in isolated contexts and exchange typed events.

Scheduling internals +

DAG scheduling supports parallel branches, cancellation, and backpressure. Context is scoped per task; typed events carry results between workers.

II. PERSISTENCE

Durable state. Resumable work.

Separate working context from durable memory. After a process restart, restore the latest checkpoint and retain completed task outputs.

Recovery semantics +

An event log records state transitions. SQLite stores local checkpoints; PostgreSQL supports shared state. Recovery replays events and retains committed outputs.

III. SOVEREIGNTY

Autonomy within explicit policy.

Enforce tool permissions, token budgets, and approval gates at the runtime boundary. Keep an audit trail for each external action.

Policy enforcement +

Scope each agent to an explicit tool allowlist, token budget, and timeout. Route privileged calls through approval gates and record their outcomes.

IV. PORTABILITY

Swap models. Keep the runtime.

Route local inference, hosted models, and existing services through a common interface. Keep orchestration independent of the inference backend.

Adapter interfaces +

Connect inference through OpenAI-compatible endpoints and tools through MCP. Switch providers in configuration without rewriting the execution graph.

02 / KERNEL ANATOMYSIX MODULES. EXPLICIT CONTRACTS.

POWER IS
IN THE KERNEL.

CONTROL PLANE / DATA PLANE

Select a module to inspect its contract.
Inputs, outputs, and execution semantics.

Skip request trace ↓

ARCHON / EXECUTION TOPOLOGY6 MODULES · 1 EVENT BUS

TYPE-SAFE EVENT BUS

goal.accepted

Runtime accepts the goal and opens an execution context.

REQUEST / #004TRACE SERVICE LATENCY

01 / ACCEPT
Runtime accepts the goal and opens an execution context.
02 / PLAN
Planner resolves the goal into a dependency graph.
03 / DISPATCH
Scheduler assigns ready branches to isolated workers.
04 / EXECUTE
Model Router and Tool Gateway return typed results.
05 / PERSIST
Runtime joins the outputs; Memory stores checkpoint #004.

DISTRIBUTED DEPLOYMENT

From localhost to the cluster.

Configure runtime →

03 / FIELD MANUAL

Runtime specification

Platforms	macOS / Linux / Windows (WSL2)
Runtime	Python 3.11+ · asyncio
Orchestration	Directed Acyclic Graph
State backend	SQLite / PostgreSQL
Tool protocol	MCP · JSON-RPC 2.0
Model interface	OpenAI-compatible / Local
Deployment	Native / Docker / Kubernetes
License	MIT

Download kernel config ↓

Boot sequence

QUICKSTART / 3 STEPS

01
Initialize a workspace
archon init my-kernel
02
Configure inference
Set the endpoint and credentials in archon.yaml.
03
Submit a workload
archon run "Trace service latency"

archon.yaml

kernel:
  name: my-kernel
  concurrency: 4
  checkpoint: true
model:
  provider: openai-compatible
  base_url: http://localhost:11434/v1
  name: qwen3
policy:
  approval: required

Deployment FAQ

How does the runtime differ from an agent framework?+

An agent framework defines reasoning and tool behavior. ARCHON manages execution lifecycles, scheduling, durable state, and authorization across agents, models, and tools.

Can inference run entirely on local infrastructure?+

Yes. Connect a local inference server through an OpenAI-compatible endpoint. Model inference, state storage, and tool execution can remain inside your network.

What happens when a worker or process stops?+

With checkpointing enabled, committed steps and execution context are persisted. Resume from the latest checkpoint. External side effects require idempotency keys or explicit reconciliation before retry.