Skip to content
Dot
DocsSign inGet started

Agents

Tools, memory, structured output and human approval.

An agent is a chatbot that can also reach for tools. It keeps everything a chatbot has — instructions, model settings, knowledge — and adds four things.

Tools

Enable tools from a registry and configure each one. When the model asks to use a tool, the request is resolved against that registry and reported in the transcript with its decision.

DecisionMeaning
SimulatedThe tool is known and enabled; the call is recorded but nothing was executed.
Needs approvalThe tool or the agent requires a person to approve the call first.
UnavailableThe tool is known but not enabled on this agent.
Unknown toolThe model asked for something that is not in the registry.

Memory

Memory bounds how much of the conversation is sent back to the model. Set a window size and, optionally, a summary of what falls out of it, which keeps long conversations affordable without losing the thread.

Structured output

Provide a JSON Schema and the agent is instructed to answer in that shape, which is what you want when the reply feeds another system rather than a person. The schema is validated as JSON when you save it; the model's reply is not yet validated against it server-side.

Human approval

Turn on approval at the agent level to mark every tool request as needing a person, regardless of the per-tool setting.