Who this is for. Site builders deciding whether an AI agent belongs in a job. Afterwards you can say what the action costs, where it may sit, and where the platform's own documentation takes over.
What it is for
An agent is for judgement a condition cannot express: classifying an enquiry by its words, drafting a reply for a person to check, summarising a long submission. It is not a way to run fixed logic, which a template does faster, at no cost, and the same way every time.
| Field (admin label) | Takes |
|---|---|
| Agent install | An agent installed for the client, from the library or custom, with its prompt and the tools it may use. Installs are managed under Agents |
| Agent input | Template code that composes the message the agent receives, from the event and earlier results. Empty sends a generic instruction |
It leaves decision (the agent's summary), result (the structured result, read by key), terminationReason (why the run of the agent ended) and invocation, the recorded invocation with its tokens and cost. A later condition branches on runData.get('classify').result.topic, for example.
Where it cannot go
- Not in a synchronous job. An agent can pause a run to ask a person and resume when they answer, which a waiting web request cannot do, so saving one on a synchronous job is refused.
- Not inside a loop. Saving it there is refused for the same reason.
While an agent waits for a person the run is waiting, and the actions after it run once the answer arrives.
The cost model
- The client's own provider key. Agents call a language model provider with a key the client enters under Agents › Providers & budgets; the platform stores it encrypted. Without a key for the agent's provider the action fails with No … API key is configured for this client.
- Paid per invocation, by the provider, to the client. A job on a busy event multiplies that cost by every record: an agent on every form submission is a cost per submission.
- Two limits. A monthly cost cap for all agents (Agents › Configuration) and a monthly budget per provider. Once either is reached, agent actions stop running until the limit is raised or the month turns.
- Re-runs. Re-running a run with an agent offers a replay of the recorded answer at no cost, or a fresh call that is charged again (Reading a job run).
Installing agents, providers and budgets, invocations and escalations are covered in the platform's own documentation: AI Agents.