Agents and skills
Agents make a working style reusable. Skills add focused procedures or domain guidance to that style. Keep the base agent small and add skills only when the task needs them.
Configure an agent
Create or edit an agent and define:
- a clear name and purpose;
- the role and durable instructions it should follow;
- a default provider and model, if appropriate;
- the tools it may use and the approvals those tools require;
- optional skills for specialized work.
Write instructions as operating rules, not a one-off task. Include how the agent should inspect context, communicate uncertainty, respect scope, and verify changes. Put the actual task in the chat.
Add a skill
A useful skill has a narrow trigger and a repeatable procedure. It should explain:
- when the skill applies;
- what inputs or access it needs;
- the sequence and checks to follow;
- what a completed result must contain.
Avoid duplicating the same rule across the agent and several skills. When instructions conflict, simplify the configuration before running it unattended.
Test before reuse
Run the agent in a small chat and verify that it:
- selects only the tools the task needs;
- asks before sensitive or irreversible actions;
- stays inside the requested project and file scope;
- produces the expected checks and delivery information;
- stops clearly when blocked.
Test a new or changed skill independently before reusing it in other chats.
Use in chats
Select an agent when starting a chat to reuse its configuration. Keep task-specific instructions in the chat. Revisit the model, skills, tool access, and limits whenever the workflow changes.
Security and portability
Agent and skill instructions can cause tools to read data or perform actions. Treat imported configurations as untrusted until reviewed. Do not embed secrets in instructions; use the credential mechanism of the provider, tool, or MCP server.
Availability can depend on the selected model and connected tools. If an agent behaves differently after a provider or model change, retest its skills and tool calls.
Next: Tools and MCP.