Skip to main content

Platform Overview

Circus is a platform for running AI-powered workflows — but the most useful way to understand it is that it's a decision instrument that happens to run workflows, not a workflow tool that happens to track cost. Its whole design turns on one capability: you change the AI model, the prompt, or the service configuration in the Circus UI, and the next run uses your new settings — without anyone editing the workflow itself. That clean separation is what lets you change one thing and read its effect on cost, quality, and outcome directly.

A mental model: your AI agents are employees​

The pieces fit together more easily if you picture your AI setup as a small organization:

  • Agents are your employees — the ones who do the work.
  • Prompts are their job descriptions — what each employee is supposed to do.
  • System context is the environment they work in — who your company is, your brand, your audience, your tone.
  • Models are their raw capability — the brain each employee thinks with.
  • Services are the outside vendors they call — a text-to-speech provider, an image generator, and so on.

You, the operator, run this organization from the Circus UI. You don't write code and you don't edit workflows; you configure the staff and their environment, and the workflows put them to work.

What you can do with it​

Circus answers three layers of question about your AI automations:

  • Is it working? Which runs succeeded, which failed, where they broke, and whether a failure was your configuration or an external service.
  • What does it cost? Spend per task, per workflow, per agent, per service — real numbers from real runs, not estimates.
  • Should we do it? With production cost and quality data in hand, you can decide honestly whether a task is worth automating, and where a human still needs to stay in the loop.

The shape of the rest of these docs​

From here, the Operator documentation follows the natural order of work: set up your account, build your configuration (context, models, agents, prompts), wire up services, engines, and workflows, then run workflows and monitor their cost and outcomes. Two ideas underpin all of it — that everything you configure is connected (so disabling one piece ripples through the rest), and that your work lives in workspaces you can branch and compare. Both are introduced where they first matter, and a full end-to-end demo at the end ties everything together.