Agents and Subagents

An agent combines instructions, tool permissions, and optional requirements for a type of work. Select an agent by the result it is designed to produce.

Primary agents

Primary agents appear in the selector below the Chat prompt. Open the selector or enter /agents, then read each agent's description before choosing it.

The available set depends on organization, global, and project configuration. Do not assume that an agent available in one project is present in another.

Changing the selected agent affects subsequent work in the session. It does not rewrite earlier messages or repeat completed actions.

Subagents

A primary agent can delegate a bounded task to a subagent. The delegated work remains connected to the parent session, and the result is returned there when complete.

The transcript shows a row for each subagent with its status. Select the row to open the child session. See Work with Subagents.

Subagents are not Agent Manager sessions. Agent Manager gives the user separate work streams and Git worktrees; a subagent performs delegated work for a parent session.

Agent requirements

An agent can require skills, MCP servers, or VS Code extensions. Reasoning checks those requirements before the work starts and identifies missing capabilities.

Install or connect the named capability, reload configuration when requested, and retry. Granting an unrelated permission does not satisfy a missing requirement.

Customize agents

Open Settings → Agents → Agents to inspect or create agents. Use a skill for reusable procedures, MCP for external tools, and a workflow for a reusable prompt.

Older versions called agents modes. That term can still appear in migrated configuration, but the current product language is agents.