Use Adapt
Agents
An agent is a named assistant with its own instructions, model, and tools, set up from the connections your workspace admins have enabled.
Every member can create agents and use any agent in the workspace. You manage the agents you created; owners and admins can manage every agent. Each agent starts from the workspace's Agent defaults and keeps its own copy, so later changes to the defaults don't change existing agents.
Create an agent
Open Agents → New agent → From scratch.
- Name the agent and pick its appearance.
- Write its Instructions: what it's for and how it should work. They're added after the company instructions on every run. See Instructions and skills.
- Under Where it answers, keep Web chat to talk to it in Adapt, and add any surfaces it should answer on, such as Slack or GitHub.
- Review its capabilities. They start from Agent defaults; change any of them for this agent.
- Choose Create agent.
If a surface you picked still needs setting up, Adapt opens the agent's setup page next. Otherwise you go straight to a chat with it.
- Name: 1–100 characters.
- Instructions: up to 16,000 characters.
- Surfaces: up to 16 per agent, including Web chat.
Choose what it can use
An agent's page has two tabs, Automations and Settings. Settings lists where the agent answers, then its capabilities. Edits stay staged until you choose Save changes.
- Models: the reasoning model the agent thinks with, and optional image and decision models. With no image model, it doesn't make images. See Models.
- Knowledge: which knowledge source it searches, or Don't use. See Knowledge.
- Sandbox: where it runs code and works with files, or no sandbox. See Sandboxes.
- Connections: the integrations it can call, such as GitHub or Linear. See Integrations.
- Limits: how much a single run may do. See Set run limits.
Two tool sets can be switched on or off:
- Harness tools: plan ahead, load skills, search the user's own past chats, and access workspace documents.
- Manage Adapt: set up agents, providers, and skills from a chat.
A capability an admin has turned off for the workspace shows Off for this workspace and can't be switched on for one agent. Choosing a connection only permits the agent to use it. A connection marked Enabled hasn't been tested; if its credentials stop working, the agent's tool calls fail and the agent page shows Needs your attention with a way to reconnect.
Set run limits
Limits apply to each run of this agent, in chats, surfaces, and automations. Leave a field blank to remove that limit.
- Steps per run: at least 1. Agent defaults: no limit.
- Time per run: 1–35,791 minutes. Agent defaults: 5 minutes. A run can still be stopped when there's no limit.
- Output per step: at least 256 tokens. Agent defaults: 4,096 tokens. Blank uses the model's own output limit.
- Context budget: at least 12,000 characters. Agent defaults: 60,000 characters. When a chat grows past it, older turns are summarized. Blank means no budget; the model's own context limit still applies.
With neither a time nor a step limit, a long task keeps running until someone stops it.
Chat with an agent
Choose Chat on the agent's page, or pick the agent when you start a chat. Only agents that answer on Web chat appear there. See Chats and tool activity.
To run an agent on a schedule or when something happens on a surface, add an automation on its Automations tab. See Automations.
Share an agent
Share → Download config saves the agent as a YAML file; Copy to clipboard copies the same text. The file holds the agent's name, instructions, settings, and automations, and refers to connections by name. It never contains connection credentials or secret values.
To bring an agent in, choose Agents → New agent → Configuration file and pick a .yaml, .yml, or .json file. A file can hold several agents. Because the file names connections rather than carrying them, the workspace you import into needs its own matching connections.
- Agent files: up to 8 MB.
Rename or delete an agent
Rename the agent or edit its appearance from its page, or from the More actions menu. Edit its instructions with View instructions.
Delete agent removes its settings and automations. Surface connections that only this agent used are disconnected too when you're allowed to delete connections. Chats with the agent stay, but can't be continued. Deleting can't be undone.