Agents are AI teammates you chat with to get work done. They live inside projects: they can use that project's files, skills, and integrations, and they show their work as artifacts next to the thread.
You don't configure infrastructure to use them. You start a thread, write what you need, and review what comes back.
Start an agent
From Home, Agents in the left icon strip, or a project's sidebar:
- Choose the project the work belongs to (or use your personal project).
- Open a new agent thread and describe the task in the first message.
- Optionally pick a model for the conversation: faster models for quick answers, default for everyday work, and deeper “thinking” models when the problem is harder. Usage for each kind shows up under Settings.
- Send. The agent streams a reply and may use tools along the way.
You can also mention @kluff in a topic or DM. In those human-first chats, the agent only runs when you mention it. In an agent thread, every message is for the agent.
Read the reply
While an agent is working you'll see it stream text and call tools (search, integrations, files, and so on). When it needs permission, an approval card appears in the thread and in Notifications.
The artifacts panel sits beside agent threads. Open it to preview what the agent produced: documents, code, images, schedules, flows, pull requests, or merge-conflict views. Artifacts are the reviewable record of “what happened, what changed, and why.”
What agents can use
What an agent can reach depends on the project:
- Files in that project, including git-backed repos. See Files & code
- Integrations enabled for the project
- Skills available in the workspace (and any you load for the task)
- Browser / computer use on the desktop app, if you turn those tools on for the project
If a tool isn't listed on the project, the agent can't use it there. That's intentional: two projects in the same workspace don't automatically share every connection.
Write a good first message
Agents do better with a concrete outcome than a vibe.
- Goal: “Draft a customer email answering these three objections.”
- Constraints: “Don't send anything. Don't change production.”
- Context: point at a file, paste a thread, or work inside the project that already has the docs.
- Tools: “Use the connected CRM to look up the account, then propose next steps.”
If the result isn't right, reply in the same thread. The agent already has the conversation; you rarely need to start over.
Safety
Agents follow workspace rules. They may ask first before using a tool, and they won't send secrets that belong in Settings. Review Approvals & notifications and Security if you're rolling agents out to a wider team.

