Agents

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

You can have as many agents as you like. One agent per job works best, for example an agent that files invoices and another that prepares your morning briefing, because the instructions stay short and specific.

Creating an agent

Open Agents in the sidebar, choose New agent, and give it a Name, a Model and its Instructions. You can change all of them later. After Create you land in the agent's admin chat.

Creating, changing and deleting agents needs the Editor or Admin role. Viewers can open agents and read their chats and settings.

Agent settings

Open an agent and choose Agent Settings. The settings panel on the right has these sections:

• General: the name and the instructions.
• Model & Capabilities: the AI model and what else the agent can do.
• Integrations: the connected accounts the agent may use, and its triggers.
• Schedules: when the agent runs on its own.
• Advanced: max steps.
• Danger zone: deleting the agent.

The chat next to the panel is the agent's admin chat, where admins can describe a change instead of making it field by field.

Instructions

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

In the instructions you can type / to reference a skill, which the agent then always has at hand, and @ to reference a file in the agent's file storage.

Model

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

A new agent starts on your organization's default model, which admins set in Organization Settings.

Capabilities

Capabilities are abilities beyond talking and using your connected accounts. Which ones you can turn on depends on the model. If you switch to a model that does not support a capability you use, the capability is turned off and you are told. A new agent starts with Web Search on when its model supports it.

• Web Search: the agent can look things up on the web.
• Code Execution: the agent can run code to calculate, convert or analyze data.
• Image Generation: the agent can create images. They are saved in its file storage, so turning this on also turns on Agent File Storage.
• Agent File Storage: a place where the agent keeps files between chats, and reads and writes them. Its settings button lets you choose which file actions the agent may take and which need your approval. See files.
• Memory: the agent remembers facts from earlier chats. See memory below.

Memory

With Memory on, facts from the agent's chats are remembered and relevant ones are brought into later chats, so the agent does not start from scratch each time.

Memory is shared across your organization: facts remembered in one chat can come up in any chat that uses memory, including other members' chats with other agents, and organization chats, which always use memory. Remembered facts cannot be viewed or edited, so leave memory off for an agent that handles information not everyone in your organization should see.

Integrations and tools

Under Integrations you choose which of your organization's connected accounts this agent may use. Switching an account on gives the agent all of its tools at once, none of them requiring approval. Open the account's settings to control:

• Which tools the agent may use. Tools that need a permission the account did not grant are greyed out; see permissions.
• Which tools require approval. With Requires Approval on, the agent pauses before using the tool, shows you exactly what it intends to do, and waits for your answer. See approving actions.

Requiring approval is the main safety control. Turn it on for anything that sends email, deletes data or spends money. It also applies to the sub-agents the agent starts and to its automated runs.

The same settings have a Triggers tab, where you set up runs that start when something happens in the account. See triggers and schedules.

Schedules

Under Schedules you can have the agent run once or repeatedly at a set time. See triggers and schedules.

Max steps

Max steps (under Advanced) limits how many steps the agent may take in a single run, where a step is one round of thinking and using tools. Raise it for long jobs; lower it to keep an agent that goes in circles from working for long. Each sub-agent gets the same limit for its own part of the work.

Deleting an agent

Under Danger zone, choose Delete agent, or use Delete in the agent's menu in the Agents list. Deleting is permanent and also deletes the agent's chats, triggers and schedules.