Agents
Agents are LLM-backed actors that run a plan→act→observe loop, call their own tools, and remember across runs.
What an agent is
An agent has a system prompt, a default model, a set of actions/skills it can call, and long-term memory. During a run it loops: read the task, decide which tool to call, observe the result, and continue until the task is done or a step budget is hit.
Anatomy
| Field | Purpose |
|---|---|
| System prompt | The agent's role, voice, and instructions. |
| Model | Which registered model it reasons with (provider-agnostic). |
| Actions & Skills | The tools it may invoke (its allow-list / wired capabilities). |
| Memory | Long-term memory is on by default; toggle per agent via memory_config.longTerm. |
| Variables | Reusable {{ name }} values injected into the system prompt. |
| Caps | Cost, time, and step budgets per run. |
Agents inside workflows
An agent node runs the agent as one step of a workflow. The agent still calls its own tools mid-run; the workflow orchestrates what happens before and after. Its reply is available downstream as ${agent_output}.
The AI builder can create agents for you — give it a workflow description and it provisions the agents with the right tools and prompts.