Agents
Your own agents for jobs you do often, each with its instructions, files and, if you like, its own model.
What agents are
An agent works like a GPT or a Gem: a name and picture, instructions, files it knows, conversation starters, what it may do, the connectors it may use, and an optional model of its own. Pick one in the message box and it answers that chat; a Cat can use one as its brain.
Make and edit them in Customize → Agents → Your agents. Earlier versions called them profiles.
Claude Code, Codex and ACP agents are on the Coding agents tab next door. See Coding agents.
Make an agent
Press New agent, or Start from a template: Research assistant, Writing coach, Code reviewer or Patient tutor. A template only fills in the editor, so change anything before you save.
| Field | What it does |
|---|---|
| Name, Picture, Description | How it shows in the agent picker. |
| Instructions | How it should behave: its role, tone, steps to follow, what to avoid. This is the most important part. |
| Conversation starters | Shown under the message box when someone picks this agent. |
| Knowledge | Files it should know. They go into its chats whole, or are searched when there are too many for the model. |
| Capabilities | What it may do. See below. |
| Connectors | All connectors, or only the ones you pick. See Connectors. |
| Model | A model of its own, or whichever model the chat uses. |
Under Advanced:
- Pick this agent for new chats: new chats start with it picked, and you can switch it off per chat.
- Run actions without asking: commands and file changes run without an approval card. Only for agents you trust with your files.
- Most steps per answer: how many tool steps it may take before it has to answer.
Capabilities
Anything switched off is not available to the agent at all.
| Capability | What it allows |
|---|---|
| Web search and browsing | Search the web and read or use web pages. |
| Files and code | Read, write and run code in a project folder. |
| Documents and images | Create documents, notes, charts and images. |
| Memory | Remember things across chats. |
| Email and feeds | Read email and RSS feeds you connected. |
| This computer and Asyncat | Use this computer (apps, clipboard, screen) and manage Asyncat. |
A new agent starts with the first four on. Planning, asking you questions and searching its own files are always allowed. Connectors are chosen separately.
Using an agent
- Pick it in the message box. Its starter prompts show on a new chat, and Start a chat in its editor opens one.
- Its memories are its own: an agent sees its own memories plus shared ones. See its memories in the editor opens them. See Memory.
- A Cat can think with it: its instructions, files and model become the Cat's brain. See Cats.
- Tools you always allow for it are listed in Settings → Privacy & safety → Always allowed, where you can take them back.
Your preferences in Settings → Personalization apply to agents too. See Personalization.
Handing work to an agent
Each agent gets a handle from its name, shown under Advanced. Mention it in any chat with #handle to hand it one message.
During a run, Asyncat's agent can also hand a task to one of your agents with the delegate_to_profile tool. The agent you named works with its own instructions, working folder, step limit and pre-approved tools, then reports back.
delegate_to_profile(
profileHandle: "code-reviewer",
task: "Review the changes in this copy for security issues"
)See Agent tools.