Agents

Agents are reusable AI worker profiles. In the UI, they let users capture how a task should be approached instead of rewriting the same system prompt, tools, permissions, skills, and routing hints for every run.

What Users Do In Agents

Open Agents from the System section of the sidebar. Create an agent manually, generate one from a prompt, edit an existing profile, attach skills or plugins, and choose whether tasks should use that agent explicitly or let routing select one.

Agent AreaProduct Meaning
IdentityName, description, key, scope, and project association help users find the right profile.
InstructionsSystem prompt and model defaults shape how the agent works.
ToolsFile, shell, web, notebook, scoped-file, and management tools define what the agent can do.
SkillsReusable instructions and slash-command style capabilities.
Plugins and MCPMarketplace-backed integrations and runtime tool servers.
RoutingDescribes when the agent is good, when to avoid it, and how confident routing should be.
PermissionsControls what workflow capabilities the agent is allowed to use.
Lifecycle hooksAdds behavior around task execution events.

Scopes

ScopeWhen To Use It
GlobalThe agent should be reusable across projects.
ProjectThe agent is specific to one repository or workspace.

How Agents Fit Tasks

When creating a task, users can choose an agent or leave selection on auto-routing. An agent can also define model behavior, so task execution can inherit a consistent combination of instructions, tools, and provider settings.

Best Practices

PageWhy It Matters
TasksTasks can run with a selected agent.
ModelsAgents can inherit or choose model behavior.
PersonalitiesPersonality settings affect tone and behavior.
WorkflowsMulti-agent workflows coordinate agents across larger work.