Skip to main content
Extensions

Custom agents

Reuse instructions, model settings, and tools for specialized engineering work.

Availability: Business plans (Teams and Enterprise). Not yet available on individual plans.
Create a custom agent for recurring work such as code reviews or implementation planning. Give it reusable instructions, a default model, permissions, and the extensions it needs, so you do not have to repeat the same requirements for every task. For a code reviewer, specify which risks to check and ask for a location, impact, evidence, and suggested fix for each finding.

Configure an agent

  1. Open Settings > Agent, choose Create Agent, and select the blank creation option.
  2. Enter a name and description that explain its role. The name cannot currently be changed after creation.
  3. Choose its default model and permissions. These apply to new tasks and collaboration runs. Changing either within a task affects only that task.
  4. Write instructions covering responsibilities, scope, and expected output.
  5. Select the required Skills, configured MCP servers, and Plugins, then choose Save Agent.
Example instructions:
Review the supplied code changes for correctness and regression risks.
Read the diff and relevant tests. Report only findings supported by evidence.
Include the location, impact, and suggested fix for each finding.
Do not modify files. Explain what additional information is needed when evidence is missing.

Choose tools

Select the Skills, configured MCP servers, and Plugins the agent needs. No Plugin is loaded when none is selected. Tool rules apply only to this agent and are saved with its configuration. Allowed tools can run without approval, denied tools are blocked, and other tools follow the default permissions.

Use an agent in collaboration

Enable Use in Issues and discussions in the agent settings. Collaboration runs execute on the current device, which must remain online. Turning this option off interrupts unfinished collaboration tasks and stops responses; existing records are preserved. Add the agent to a discussion, then use @ to request analysis, comparisons, or test suggestions. For work involving several roles, include it in a custom agent team. In a discussion, only an agent’s owner can invoke it with @. You can only @ your own agents, not another member’s agents.

Update an agent

Use the edit button in the agent list to update its description, instructions, and execution settings. Turning off Available for assignment interrupts all unfinished tasks and stops new tasks; historical records remain.