Agents

An agent is a teammate with one job. It has its own chat, a browser window and folder on your crew's computer, a memory, its own model and its own routines.

Talking to an agent

Write what you want in plain words. The agent works step by step and shows each step on a working card. For bigger tasks it shows a plan and ticks steps off as it goes. Every reply shows which model wrote it.

Stop: while an agent works, the send button turns into a stop button. Nothing else runs after you stop it. Reply: hover over any message and press Reply, so the agent knows exactly which message you mean. Messages can't be edited or deleted, so the chat stays a true record of what was asked and done.

Approvals and rules

Risky actions can wait for you as an approval card: sending messages, spending money, deleting, posting publicly, or running a command that changes things. You choose which in Settings → Approvals. Auto-review lets a quick AI check allow low-risk actions and asks you about the rest. Agents can still make mistakes, so check anything important. Read-only work, like reading a page or listing files, runs without asking.

• Approve runs it once. Always allow stops asking for that kind of action from that agent. Deny tells the agent no, and it tries another way.
• To make agents always ask before sending, spending, deleting or posting, set that action to Ask every time in Settings → Approvals.
• Tell an agent something like “always ask me before deleting anything” and it becomes a rule, listed in its Details. Larkgrass holds the agent to it.

Memory

When you tell an agent something lasting (“sign off as Priya, no emoji”), it saves it and shows Updated memory. It reads its memory at the start of every task. See or remove anything under Details → What it remembers. When a fact changes, the newer one replaces the old.

Team memory (Settings → Team memory) is known by every agent. Long chats are summarised as they grow, and agents can search past chats when you ask about something from earlier.

Models

Each agent can use a different AI model. New agents start on the default, DeepSeek V4.1 Flash. Pick another in the agent's Details → Model. Pro and Max plans add the frontier models (Claude, GPT and Gemini). You can also add any model with your own API key in Settings → Models; models on your own key don't count toward your AI usage (their computer time still counts).

Models marked Catalog or Small model aren't among our recommended picks and can be less reliable on long, multi-step tasks.

Editing and removing agents

Details → Edit agent changes an agent's name, job, shape and colour. You can also just ask: “Rename yourself to Atlas.” To remove an agent, open its Details and press Delete agent at the bottom.