Skip to main content

The workflow_config Snapshot

The workflow_config_snapshot holds the operator's agent configuration for this run: which agents are available and, for each, the model, prompt, and parameters it should use. In practice you rarely read it by hand — the Agent node reads it for you, matching the agent slug you give it — but understanding its shape makes the Agent node's behaviour obvious.

Shape​

The snapshot names the configuration set and lists agent assignments — one per agent the operator wired in:

{
  "workflow_config_id": 7,
  "workflow_config_name": "Script Generation Config A",
  "agent_assignments": [
    {
      "agent_slug": "script_writer",
      "model_name": "GPT-4o",
      "model_provider": "openai",
      "model_base_url": "https://api.openai.com/v1/chat/completions",
      "prompt_slug": "generate_script",
      "prompt_text": "You are a script writer...",
      "temperature": 0.7,
      "max_tokens": 4096
    }
  ]
}

Each assignment carries everything needed to make one AI call: the model identity and its base URL (the endpoint), the prompt text (the active version, resolved for you), the provider (so the right credential and request shape are chosen), and the temperature and max_tokens.

How the Agent node uses it​

When you drop an Agent node and set its agent slug, the node finds the assignment whose agent_slug matches and uses that entry's model, prompt, and parameters — nothing is hardcoded in n8n. This is the whole mechanism behind operator control: the operator swaps a model or edits a prompt in the Circus UI, that changes the assignment, and your next run's snapshot carries the new values automatically.

If the slug you configured isn't in the snapshot, the Agent node treats it as a fatal misconfiguration — it logs the problem, terminates the run, and errors — because it can't guess which agent you meant. So the slug on your Agent node must match one the operator actually configured.

The snapshot carries the model identity and consumption units, but not pricing. Cost is computed by the platform from current registry pricing when a step is logged — you report tokens, the platform prices them.