Skip to main content

The Agent Node

The Agent node is the heart of AI work in Circus, and the one node that most embodies the platform's promise. Its whole reason to exist is this: you place it and say which agent it should act as; the operator controls everything else. Change a model from GPT to Claude, or rewrite a prompt — the operator does that in the Circus UI, and your next run just uses it. No n8n edit, no hardcoded provider config.

What it does, end to end​

When an Agent node runs, it:

  1. Resolves its configuration from the snapshot — it looks up the agent you named (by slug) in the workflow_config_snapshot and reads that assignment's model, prompt, provider, and parameters.
  2. Builds the prompt — the agent's assigned prompt as the instruction, your runtime input as the work item, and (optionally) system context.
  3. Calls the AI provider — constructing the right request for that provider and authenticating with your provider credential.
  4. Logs the result — reporting token usage, duration, and the full request/response to the platform, which prices it.
  5. Checks the run's limits — and self-terminates if the run has breached its cost or time budget.

Each of those is covered in its own article in this section; this page is the map.

The one thing you configure: the agent slug​

From your side, the essential input is the agent slug — which of the operator's configured agents this node should be. That, plus an input (the work item), is usually all you provide. Everything that makes the call work — the model, its endpoint, the prompt text, the temperature, the token limit — comes from the snapshot, not from you.

Why "no hardcoding" is the whole point​

It's worth sitting with why this matters. If you hardcoded a model and prompt into n8n, then every time the operator wanted a cheaper model or a better prompt they'd need you to edit the workflow — and the platform's core value (change one variable, measure the effect) would evaporate. The Agent node exists precisely so the operator can iterate freely while your pipeline stays untouched. Build your AI steps around it, resist the urge to bake in a model or prompt, and you've done your part of the deal.