Automate a recurring job | Sovereign AI Docs

Flows turn a job you keep doing by hand into a governed automation with human checkpoints and a run history you can trust.

Build it on the canvas

Open Flows and create a new flow. You get a canvas, a node palette, and an AI Flow Editor: describe the job in a sentence and it drafts the graph, or ask it for a change to edit the current graph. Build by hand from the palette when you want control: Crew and Swarm nodes launch multi-agent teams as a step, Agent runs a single role, Set Variable and Echo manage run state and notifications, HTTP Request and GitHub reach external systems, Decision branches on state, Loop repeats over items, Human pauses for sign-off, Report compiles the deliverable.

The Weekly Vendor Risk Review flow: schedule trigger, HTTP fetch, threshold variable, notify, scoring loop with an agent body, decision branches to a swarm deep-dive or a cleared log, CISO approval gate, report and final notification — ↳A real pipeline: fetch vendors, score them in a loop, deep-dive the risky ones with a swarm, gate criticals behind the CISO, file the report.

The node palette open beside the pipeline: Crew, Agent, Set Variable, Decision, Echo, HTTP Request, Loop, Human, Report, Swarm, Trigger — ↳Each node is a control point too. Human pauses for a signature, Decision stops a bad branch from running unsupervised.

Keep a human in the loop

Drop a Human node anywhere a person should sign off. The run pauses at that node, the run badge flips to Awaiting approval, and everything downstream stays queued until someone decides. Decision nodes route without waiting: point one at a state key with an operator (equals, greater than, present, contains) and it branches on what the run actually produced.

A live run paused at the Human approval node: the review brief, Approve and Reject controls with a required approval note, and the report queued behind the gate — ↳The run waits at a named gate: approve with a note on the record, or reject with direction. The report stays queued until a person decides.

Trigger it

1Run the flow manually while you are building; every run is inspectable.
2When it works, set the trigger to Schedule so it runs itself on a cadence.
3Or set it to Event and wire an event source, such as GitHub activity arriving through your organization's connected apps.

Trust it

Every run records durable trace events: which node ran, what it produced, what failed and why. "What actually happened" is a record you open, not a memory. The HTTP response bytes, the agent's output, the decision's taken branch, and who approved the gate all sit in the trace, and the report node compiles a per-step receipt into the deliverable itself.

The completed run: Human approval succeeded with the approval note on the record, the report compiled with a per-node receipt, the final notification fired — ↳The receipt: who approved, what every node did, and the report that proves it.

Flows run under the same governance as everything else: content policy applies to agent nodes, and human nodes make approval part of the automation itself.