Copilot Custom Agents Pack: 12 .agent.md Agents + Checker

Every agent uses only the fields that GitHub Copilot reads everywhere (name, description, tools with the standard aliases), so the same files work in Copilot in VS Code, Copilot CLI, JetBrains/Eclipse/Xcode and the Copilot cloud agent on GitHub.com. Read-only agents have no edit or execute tool, so they cannot change your code.

python3 tools/agentkit.py list
python3 tools/agentkit.py install /path/to/repo                    # -> .github/agents/ (commit it to share with the team)
python3 tools/agentkit.py install /path/to/repo --agent reviewer --agent planner
python3 tools/agentkit.py install --user                          # -> ~/.copilot/agents/ (just for you)

Existing files are kept unless you pass --force (which saves a .bak). --dry-run shows what would be written. Then pick the agent from the agents dropdown in Copilot Chat, or assign it to a task on GitHub.com.

python3 tools/agentkit.py check .github/agents

Flags what makes an agent silently misbehave:

• missing description (required), YAML errors, unknown fields, retired infer
• old .chatmode.md files that need renaming to .agent.md
• handoffs, agents, argument-hint, advancedOptions: read only by the VS Code Local harness (deprecated, removed February 2027) and ignored by Copilot
• tool names Copilot ignores (for example old Local names like codebase), tools: []
• model-policy: required without a model, an agent nobody can invoke (user-invocable: false + disable-model-invocation: true), invalid target
• mcp-servers without command/url, or with literal secrets in env instead of ${{ secrets.COPILOT_MCP_... }}
• prompt bodies over the 30,000-character limit, #file:/#tool: references, broken Markdown links, secrets in the file, duplicate agent ids

Exit code 1 on errors; --json for CI.

python3 tools/agentkit.py to-claude .github/agents --out .claude/agents

Writes Claude Code subagents: name in kebab-case, description, and tools mapped (read -> Read; search -> Grep, Glob; edit -> Edit, Write; execute -> Bash; web -> WebFetch, WebSearch).

./tools/demo.sh lists the agents, installs them into a temp repo, checks them (0 errors), checks examples/broken-agents (3 files with 13 typical mistakes), converts to Claude Code subagents.

./run_tests.sh (12 tests, Python 3.9+): all 12 agents pass the checker with zero warnings and use only portable fields; read-only agents have no edit/execute tools; each broken example is caught; secrets, size limit, empty body, YAML errors, duplicates; valid MCP secret references and tool forms pass; install is idempotent with --force backups; Claude conversion; the demo runs.

• Agents guide Copilot's behaviour; they don't guarantee output. Review what an agent changes.
• Tool availability depends on the Copilot client and your MCP setup; the checker validates names, it cannot confirm a tool exists at runtime.
• Copilot's custom agent format is evolving; written against the GitHub and VS Code custom agent references as of November 2026.