VS Code Copilot Custom Agent Creator

Trigger Scenarios

Use when the user needs to create a custom Copilot Agent in VS Code, configure a specialized AI role, define handoff workflows, set tool permissions, or write agent instructions. Trigger keywords include "create agent", "custom agent", "VS Code agent", "copilot agent", "agent.md", "handoff", etc.

⚡ Workflow

1. Understand requirements → Ask user about: agent role, available tools, whether handoff chain is needed
2. Generate file → Create .agent.md file following the template
3. Placement → Tell user to place it in .github/agents/ (workspace) or ~/.copilot/agents/ (user-level)

File Locations

Workspace agents can have additional search paths configured via chat.agentFilesLocations

File Structure

---
name: Agent Name                    # Optional, defaults to filename
description: Brief description      # Displayed in input placeholder
argument-hint: Argument hint        # Optional, guides user input
model: GPT-5.2 (copilot)            # Optional, string or array (attempt in order)
tools:                              # Tool list
  - search/codebase
  - web/fetch
  - edit
agents:                             # Available sub-agents, * for all, [] to prohibit
  - Researcher
user-invocable: true                # Optional, whether to appear in dropdown (default true)
disable-model-invocation: false     # Optional, prevent being called as sub-agent by other agents
handoffs:                           # Optional, handoff chain
  - label: Button text
    agent: Target agent
    prompt: Pre-filled prompt
    send: false                     # Optional, whether to auto-send
    model: GPT-5.2 (copilot)        # Optional
hooks:                              # Optional (Preview), scoped hooks
  PostToolUse:
    - type: command
      command: "./scripts/format.sh"
---

# Agent Instruction Body

Write the agent's behavior guidelines, workflow, output format, etc. here.
Supports Markdown links to other files.
Reference tools with: #tool:web/fetch

Frontmatter Field Reference

Handoff Workflow

Handoffs display buttons after the agent responds, allowing one-click switching with pre-filled context:

Plan → Implementation → Code Review

Example:

handoffs:
  - label: Start Implementation
    agent: implementer
    prompt: Implement the plan above.
    send: false
  - label: Review Code
    agent: reviewer
    prompt: Review the changes for security and quality.

Built-in Tool Reference

Reference tools using #tool:<name> syntax, e.g., #tool:web/fetch

Common Agent Templates

Planner Agent

---
name: Planner
description: Generate implementation plans
tools: ['search/codebase', 'web/fetch', 'search/usages']
model: ['Claude Opus 4.5', 'GPT-5.2']
handoffs:
  - label: Implement
    agent: implementer
    prompt: Implement the plan above.
---
You are in planning mode. Generate a detailed plan with:
- Overview, Requirements, Implementation Steps, Testing.
Do NOT make code edits.

Code Reviewer Agent

---
name: Reviewer
description: Review code for quality and security
tools: ['search/codebase', 'web/fetch']
---
Review code changes for:
- Security vulnerabilities
- Code quality issues
- Performance concerns
- Adherence to project conventions

Implementer Agent

---
name: Implementer
description: Implement code changes
tools: ['edit', 'read/terminalLastCommand', 'search/codebase']
---
Implement changes following existing code patterns.
Make minimal, focused edits. Run tests after changes.

Creation Process

1. In Chat, type /agents or click the gear icon → Agent Customizations editor
2. Select New Agent (Workspace) or New Agent (User)
3. Enter filename → .agent.md template generated
4. Fill in frontmatter + instruction body
5. Or use /create-agent to have AI generate based on description

Claude Format Compatibility

.claude/agents/*.md supports Claude-specific fields:

• tools: Comma-separated string ("Read, Grep, Glob, Bash")
• disallowedTools: List of prohibited tools

VS Code automatically maps Claude tool names to VS Code tools.

Organization-level Agents

Enable github.copilot.chat.organizationCustomAgents.enabled: true to automatically discover organization-level agents.

See: references/agent-format.md