Skip to main content

Workflow Development: Overview

Welcome — you're here to build the workflows that do the actual work in Circus. Before any nodes, it's worth being clear about what a Circus workflow is, because it shapes every decision you'll make.

You build the pipeline; the operator tunes the intelligence​

In Circus there's a deliberate division of labour. You build the pipeline — the shape of the work, the steps, the branches, the error handling. The operator decides the intelligence — which model each step uses, what the prompt says, what it costs, what limits apply — and they do that from the Circus UI, without ever touching your workflow. That's the platform's core promise, and it has a direct consequence for you: don't hardcode what the operator is meant to control. No hardcoded models, no hardcoded prompt text. You'll see exactly how that works when we reach the Agent node and snapshots.

The workflow is an accountable team member​

A Circus workflow isn't a black box that quietly does its thing. It's a member of the team, and it's expected to be accountable — to report what it did, what it consumed, and how it ended — so the organization can see cost, quality, and outcomes with precision. That accountability isn't extra homework you bolt on; it's what makes your work matter. The logs and costs your workflow reports are exactly the production data the business uses to answer "is this working, what does it cost, should we keep doing it." Build it well and your workflow becomes trustworthy; build it as a black box and it's just a demo.

The five nodes are how it earns that trust​

Circus gives you five custom n8n nodes, and the best way to see them is as the tools that make a workflow a good citizen:

  • Init establishes the shared execution context every other Circus node relies on.
  • Agent runs operator-controlled AI — reading the model and prompt from the run's snapshot, so nothing is hardcoded.
  • Log records a step's work and consumption, and checks the run against its limits.
  • Complete ends the happy path and hands back the results.
  • Terminate ends an error branch cleanly.

Along the way they handle the things that separate production-grade from proof-of-concept: cost and time limits, traceability, and staying in sync with the platform. None of it costs you much effort — the nodes do the heavy lifting — but understanding why each one matters is what the rest of this package is for.

What's ahead​

From here: get set up (install the nodes, learn what each does), stand up a workflow's entry point (The First Steps), learn to read what the platform hands you (snapshots), do the work (the Agent node and service calls), report back honestly (logging, cost, sync), and finish cleanly (Complete and Terminate). A pointer to the full demo lets you watch it all run end to end.