Create an agent

A native Redcast agent is a profile for doing work in a folder. It uses the model selected in the chat, plus its own instructions, tools, skills, memory and permission rules. Create one in Settings → AI → Agents.

Your first agentCopy Anchor Link

1. Enable AI and connect a model that supports tool calling.
2. Create an agent, give it a recognizable name and choose a working folder.
3. Keep Runs on set to the chat's model for a native agent. Choose an ACP runtime only if you want an installed coding agent to run the task.
4. Select the tool families it needs. Start with Files for a document-reading agent; add Terminal only for tasks that need commands.
5. Leave change and command permissions at their asking defaults while trying it.
6. Open a chat with the agent and ask it to list the folder's top-level files.

The working folder is context for paths and project instructions, not an operating-system sandbox. Read the tool and permission settings before giving an agent access to personal data or commands.

Give it instructionsCopy Anchor Link

The agent reads its folder's AGENTS.md, enabled skills and its own notes. Use the editor's instructions for the agent's role and the project's AGENTS.md for rules that should follow the folder.

Work on the documentation in this folder.
Read the existing guide before changing its terminology.
When asked to edit, make the smallest useful change and report the affected paths.

A small, concrete instruction is easier to follow than a long list of aspirations. Describe the expected output and the checks that matter for your project.

Tools and permission promptsCopy Anchor Link

File reads and the agent's own memory do not need the same permission as file changes or shell commands. The default asks before changes and commands. A dialog can allow the action, allow that family for this chat, or decline it.

Allow for This Chat applies to the specific family and risk. It does not turn every tool into an unrestricted operation. Switching the chat's agent clears those temporary grants.

The editor's Permissions rows can choose Ask Each Chat, Always Allow or Never Allow for the available risks. Optional Notes additions and Screen reads retain their own asking rules. See tools and limits.

Choose the agent automaticallyCopy Anchor Link

Agent for new chats sets the agent used by new AI Chat conversations, including an ACP agent. You can still switch it in the composer. An empty new chat does not start an ACP process merely because the default is shown.

A folder can set its own agent for New Chat in Folder. Quick AI has its own setting, Quick AI answers with, which accepts native agents. Existing conversations keep their own agent.

Memory and skillsCopy Anchor Link

An agent can keep its own notes through its Memory tools. These are separate from your Notes feature. Skills are named sets of instructions it can load when relevant; choose which ones it can see in the skills editor.

Continue without the windowCopy Anchor Link

Type @handle and a task in the launcher to run the agent in the background. Its result becomes an ordinary chat. For repeated tasks, create a schedule.

What to check when a task failsCopy Anchor Link

Open its tool rows: they show the operation and arguments, such as the command being run. A refusal usually identifies the permission or scope that blocked it. A provider error may mean the selected model does not implement tools, has no balance, or has reached a limit. Change one thing at a time and retry with a small task.