Agents

An agent is a reusable AI worker that belongs to your organization. It has instructions describing what it does, a model, a set of capabilities, the integrations it may use, and its own chat history.

You can have as many agents as you like. In practice one agent per job works best — an agent that files invoices, an agent that prepares your morning briefing — because the instructions stay short and specific.

Instructions

Instructions are the agent's brief: what it does, how it should behave, what to do in edge cases. Write them as you would explain the job to a new colleague.

In the instructions editor you can type / to reference a skill and @ to reference a file in the agent's storage. A skill referenced here is always available in that agent's chats.

Model

Each agent has a default model, used for its chats unless you pick a different one in a particular chat. The picker shows each model's input and output price per million tokens, so you can weigh capability against cost. See billing and credits.

Capabilities

Capabilities are extra abilities beyond talking and using your connected tools. Which ones you can enable depends on the model — if you switch to a model that does not support a capability, it is turned off automatically and you are told.

• Web search — the agent can look things up on the web.
• Code execution — the agent can run code in a sandbox to compute, transform or analyse data.
• Image generation — the agent can create images. This needs file storage, which is enabled automatically.
• Agent file storage — a persistent workspace where the agent can keep files between runs, and read and write them. You can browse it under Files.
• Memory — the agent remembers facts about you across chats, rather than starting fresh each time.

Integrations and tools

Under Integrations in the agent's settings you choose which of your organization's connected accounts this agent may use. For each one you control:

• Which tools are enabled.
• Which tools require approval — the agent pauses and asks before running them, and you see exactly what it intends to do.

Requiring approval is the main safety control. It is worth turning on for anything that sends email, deletes data, or spends money.

Schedules and triggers

An agent can start work without you: on a schedule, or when something happens in a connected account. Both are described in triggers and schedules.

Advanced

Max steps limits how many tool-loop iterations the agent may take in a single run — between 1 and 200, and 100 by default. Raise it for long multi-step jobs; lower it to keep a misbehaving agent from working for a long time.

Deleting an agent

Deleting an agent is permanent, and it also deletes that agent's chats, triggers and schedules.