Agents and skills
Agents define who performs work. Skills provide reusable instructions that agents can load when the work calls for them. Both are created from Settings and belong to the scope selected at the top of the Settings sidebar.
Create or edit an agent
- Open Settings and choose Workspace, a Team workspace, or a project in the scope selector.
- Open Agents. Search or select an existing agent to edit it, or choose New Agent.
- For a new agent, choose Start blank or Copy of existing. Copying is useful when a project needs its own version of a workspace agent.
- Set the name, brief description, and prompt. Choose a tier, and choose a provider only when this agent should prefer one over the default provider.
- Set the fence for actions outside the agent's project boundary, optionally choose an icon, and save.
The available tiers and their current model assignments come from Workspace > Providers. Keep the agent at Any backend when the workspace's default provider should decide. A provider preference narrows that choice to the selected provider. See Providers and backends.
An agent created in the personal workspace is available across projects. An agent created in a project belongs to that project. The list can show inherited workspace agents alongside project agents. To customize an inherited agent for one project, use Copy to project to edit or create a copy from New Agent; the creation dialog warns when the project copy will shadow an inherited agent with the same name.
Create or edit a skill
- Select the workspace, team workspace, or project that should own the skill.
- Open Skills, then choose New Skill or select an existing skill.
- Start blank or copy an existing skill into the selected scope.
- Enter its name, description, and instructions. Enable Restrict allowed tools only when the skill should limit the tools available while it is active.
- Save the skill.
Skills at the personal workspace are available across projects; project skills belong to that project; team-workspace skills belong to that team's shared configuration. The current Agent editor does not include a per-agent skill-assignment control. Agents receive the skills available in their execution context and can load a relevant skill when its description matches the work. A person can also invoke an available skill from chat by typing its slash command.
Use specific descriptions. The description is what lets an agent recognize when the skill applies; the instruction body carries the procedure it should follow.
Advanced packages and integrations
The Settings editors cover the normal agent and skill workflow. Skill scripts, package files, external Model Context Protocol servers, and workflow code are extension mechanisms. See Scripting and workflows for those surfaces instead of editing package internals for routine configuration.