# Agent factories (/docs/developers/agent-factories)



ox works with any tool that orchestrates coding agents — OpenClaw, Factory Droid, custom orchestration. However the session is launched, the integration is identical: the agent primes on startup, Team Context flows in, and artifacts are captured back.

The integration is the same [#the-integration-is-the-same]

An orchestrator changes how a session starts, not what ox does inside it. The orchestrator spawns the agent, the agent runs `ox agent prime`, and from there the [prime-work-capture loop](/docs/developers/how-it-works) runs unchanged.

<Mermaid
  chart="graph LR
    O[&#x22;Orchestrator&#x22;] --> A[&#x22;Agent starts&#x22;]
    A --> P[&#x22;ox agent prime&#x22;]
    P --> W[&#x22;Agent works&#x22;]
    W --> C[&#x22;Artifacts captured to Ledger&#x22;]"
/>

Because priming is wired into the repo's `CLAUDE.md` (or the equivalent for other agents), it fires regardless of launch path — terminal, IDE, API, or factory. You don't configure the orchestrator to know about ox; the agent carries the integration with it.

Many agents, one Team Context [#many-agents-one-team-context]

When a factory runs several agents in parallel against the same repo, they all prime from the same [Team Context](/docs/features/team-context). Every agent sees the same conventions, decisions, and domain terms — so parallel work stays coherent instead of drifting apart.

<Mermaid
  chart="graph LR
    T[&#x22;Team Context&#x22;] --> A1[&#x22;Agent 1&#x22;]
    T --> A2[&#x22;Agent 2&#x22;]
    T --> A3[&#x22;Agent 3&#x22;]
    A1 --> N[&#x22;New decisions&#x22;]
    A2 --> N
    A3 --> N
    N --> T"
/>

As each agent's session is captured, its decisions feed back into the shared knowledge that future sessions consume. This is the multiplayer loop at scale: more agents working in parallel means the context compounds faster.

<Callout type="info">
  Captured artifacts become available to other sessions once committed. Agents running concurrently see each other's work on the next prime, not mid-session — so coordinate parallel agents around natural commit boundaries.
</Callout>

What's next [#whats-next]

* [How it works](/docs/developers/how-it-works) — the loop every orchestrated session runs
* [Session recording](/docs/developers/session-recording) — how captured work becomes shared memory
* [Coding agents](/docs/developers/coding-agents) — which agents orchestrators can drive