What an agent can do

An agent answers, acts, and hands off — inside the rules you author. It steers its own behavior with directives, answers from your data, runs multi-step routines across turns, takes action through your tools, and gets a person when it should. Every part of it is recorded in the turn trace.

This page explains the four capabilities and how they fit together. For the shape they add up to, see Guided autonomy.

Steer behavior with directives

A directive is a standing rule. It pairs a condition with how the agent should behave when that condition holds.

When the customer sounds anxious, slow down and confirm before acting.

Directives are always on, and they steer the reply itself: its tone, its judgment, its approach. Giving the agent a new thing to do is a skill’s job. Each turn, the matching directives are applied by meaning rather than by a keyword list, so the same rule works across every language and surface.

Every directive that applies is recorded in the turn trace, with the reason it matched.

Answer from your data

When the agent has evidence in your documents, it answers and cites the source. When it does not, it says so. See Grounded answers for how grounding works and how to verify it.

Run multi-step routines

A routine is a flow the agent runs across turns. It can collect the values it needs, call a skill, branch on what happened, and finish or hand off to a person.

You author a routine in plain language in the agent’s Routines settings. Write the steps the way you would explain them to a teammate. Where a step needs structure, drop in a chip:

• a variable to collect (@email, @order_id), stored as a typed value
• a skill to call as a tool step
• a condition to branch on, decided in code or by the model
• an end or a handoff to a person

The platform compiles that draft into a graph the engine runs and resumes turn to turn. You validate, then publish, and new conversations pick up the published version straight away. Conversations already running stay on the version they started on.

A simple example — a return request:

1. Collect the order id.
2. Check how old the order is.
3. If it is within the return window, start the refund. Otherwise, hand off to a person.

To author one, see the in-app Routines editor, or the authoring API under /api/v1/agents/{agentId}/routines (Agents and skills).

Take action and hand off

A routine is where talking turns into doing:

• Call a tool. A tool step dispatches a skill — for example, a skill backed by an external system over MCP — and the routine branches on the result.
• Fire a webhook on completion. A routine can post a signed payload to a workspace webhook destination when it finishes, so a downstream system picks up the result.
• Hand off to a person. A handoff ends the routine by escalating to a human. The built-in contact flow — collect an email, collect a message, submit, confirm — is written as routine data too, registered with the platform rather than stored on your agent.

These run through an asynchronous action outbox: actions are dispatched, retried on transient failure, and recorded.

See why it did what it did

Directives, skills, and routines share one debug surface. Each directive match, each skill dispatch, and each routine step is recorded in the turn trace with the reason it applied. If the agent does the wrong thing, you can see which rule steered it, which tool it called, and which step it was on — the same way you inspect a grounded answer.