Durable Task Delegation

A normal tool call is one-shot: sportsclaw asks, the tool answers, the turn ends. Some work is bigger than that — research that spans many steps, jobs that wait on async data or human input, tasks that should survive a restart. For those, sportsclaw can hand the work to the Machina durable loop running on a connected pod — it dispatches the long-running work and reads the result back when it's ready.

How it works

When you connect a Machina pod that runs the durable loop (the loop-runner agent — see Machina), sportsclaw automatically exposes a machina_loop tool. The loop lives on the pod, not in sportsclaw: every turn is persisted as a document and resumed by the pod's beat, so it survives interruptions, async tools, and waiting on input. sportsclaw's own loop stays fast and ephemeral — it delegates the long-running work, then reads the result back.

The agent drives three actions:

Action	What it does
start	Begin a new durable session for a task. Returns a session_id.
continue	Add a follow-up message to an existing session.
read	Fetch the current state — latest reply, status, turn count.

You don't call these yourself — the agent does, when it decides a task is better run durably. You just ask:

sportsclaw "Track this week's match series and compile a pre-game analysis for each game"

Starting or continuing a durable session is gated by the same approval prompt as other side-effecting actions, so you stay in control of what gets delegated. The pod's reply is treated as untrusted external data on the way back in.

Requirements

• A connected Machina pod that runs the durable loop — wire one up with sportsclaw machina connect (or sportsclaw mcp add). The machina_loop tool only appears when such a pod is connected.
• The pod must expose the loop's execute_agent and search_documents tools over MCP.

Check what's connected with sportsclaw doctor (it lists Machina pods) and sportsclaw mcp list.

Ephemeral vs durable

Use direct tools for a quick, in-the-moment answer. Reach for the durable loop when the work is long, multi-step, or needs to outlive a single exchange.