Flows & capture

A flow is an ordered sequence your agent walks — step by step, reasoning through each one, and writing what it finds back into the brain. Instead of handing an agent everything at once in a prompt, you hand it one thing at a time, in the order you set.

How a flow runs

1. The agent claims the next step.
2. It does the work for that step.
3. It writes the result back into the workspace (a doc, a task, a decision).
4. It moves to the next step — or logs a blocker.

Because every step writes back, the work accumulates in the graph and every other client sees it live.

Capture

A capture node lets an agent write a document at walk-time — it produces a finding and files it directly, no copy-paste. Findings from a flow become part of the same brain everything else reads.

Running a flow

Start a flow run from the app, or over MCP with the flow tools (list_flows, get_flow, start_flow_run, submit_flow_step_result, submit_flow_capture). See the API reference for the full tool list.