#Agents

An agent is an AI model with a job, a set of tools and optional knowledge, defined in the designer and run by the same engine as every process.

#An agent is a process

Agents are not a separate product. An agent is created, validated, deployed and monitored exactly like a process, and every rule about deployments and runs applies to it unchanged.

flowchart LR
  subgraph Agent
    M["Model and instructions"]
    T["Tasks"]
    Tools["Tools: adapter actions, flows, MCP servers"]
    K["Knowledge vaults"]
  end
  Chat["Chat in Execution"] --> Agent
  CA["Call activity in a process"] --> Agent
  In["Inbound adapter"] --> Agent
  Agent --> R["Result back to the caller"]

#What you configure

Part	What it decides
Model	Which model the agent uses, its temperature and limits, and its instructions. The model is reached through an LLM provider set up under Systems. See Configure an agent's model.
Tasks	The jobs the agent can be asked to do, each with its own prompt, input and expected result. A task's prompt can pin notes, folders and tools with `@` — see Pin notes, folders and tools into a task prompt. See also Let an agent call its own tasks. See Define what a task takes and returns below.
Tools	What the agent may do while it works: call adapter actions of your systems, run flows, use MCP servers. See Give an agent tools.
Knowledge	Notes the agent reads — and, if allowed, writes. See Agent knowledge vaults.

#Define what a task takes and returns

A process that calls an agent calls one of its tasks. Each task is listed on the agent's Tasks tab. Expand a task to see its instruction at the top and its contract below it: the input fields a caller passes in and the response fields the agent returns. Both sides may stay empty. With no response fields, the caller gets the agent's text.

To change a task, click Edit on the agent. The expanded task then becomes editable in place. Write the task's instruction in the same formatted-text editor as the agent's prompt; it is added to that prompt for this task only. The instruction box keeps one fixed height and scrolls inside, so a long instruction never pushes the contract out of view. Add, rename or remove input and response fields. Every change counts as an unsaved change of the agent. Save stores it with the rest of the agent, and Discard undoes it. A field without a name is marked, and the agent cannot be saved until you name the field or remove it.

If a stored contract uses something the field editor cannot show, that side stays locked, and its original text is kept unchanged. To define that side from scratch, discard the stored contract in the notice above the side.

A task is also a tool the agent itself can use: see Let an agent call its own tasks.

#How an agent is started

In a chat — people talk to it under Agents in the Execution app.
From a process — a call activity runs the agent and waits for its result.
From an inbound adapter — an incoming event starts it, like a message start event.

Each of them runs the agent's production deployment. See Start an agent.

#Follow what it did

Every agent run is an instance like any other. Monitoring shows its thought loop: every tool it called, with what, and what came back. See Follow what an agent did in a run.