What is an Agent? Two Formats, One Concept

Two Agent Formats

• An agent is a persona — model + tools + constraints. Same concept in both formats, different capabilities and frontmatter fields.
• Claude Code plugin agents (`agents/*.md`) — deep platform integration
  • › `model:` — explicit tier (Opus/Sonnet/Haiku)
  • › `permissionMode:` — plan (read-only) or acceptEdits
  • › `isolation: worktree` — agent gets its own repo copy
  • › `memory:` — persistent across sessions
  • › `maxTurns:` — safety cap
• Copilot agents (`.github/agents/*.agent.md`) — multi-model and interactive
  • › `tools:` — read, edit, search, execute, codebase
  • › `handoffs:` — interactive buttons to chain agents (VS Code)
  • › `mcp-servers:` — inline MCP config in YAML
  • › `model:` — fallback chain: `[claude-sonnet, gpt-4o]`
  • › `allowed-tools:` — auto-approve these tools

Choosing & Creating Agents

1. 1 We maintain both — 13 plugin agents for Claude Code, 25 Copilot agents in `.github/agents/`. Generated from shared templates by the init script.
2. 2 Tool name differences between platforms
   • › Claude Code: `Read, Glob, Bash, Edit`
   • › Copilot: `read, codebase, execute, edit`
   • › Templates include both — unrecognized names are ignored per platform
3. 3 Know which format to use
   • › Need model tiering or worktree isolation — Claude Code plugin agent
   • › Need handoffs or multi-model fallback — Copilot .agent.md
4. 4 When adding a new agent, create both formats from shared templates so all platforms are covered