Agents — Examples for GitHub Copilot
Agents are specialised AI personas you can invoke in GitHub Copilot Chat. Each agent has a defined role, expertise, and workflow that shapes how Copilot responds.
How agents work
When you select an agent in Copilot Chat (using @agent-name), Copilot adopts that persona for the conversation. The agent’s instructions override the default behaviour, giving you focused, role-specific assistance.
Agent files live in .github/agents/ and use the .agent.md extension.
How to use these examples
- Copy the
.agent.mdfile into your repository’s.github/agents/directory - Edit any project-specific details (service name, tech stack, team conventions)
- Open Copilot Chat and select the agent from the agent picker
Keep agents current. Defra standards evolve. Revisit agent files quarterly or when the Defra software development standards (opens in new tab) repository is updated.
Available examples
| Agent | Purpose | Best for |
|---|---|---|
| Defra App Developer | Builds Defra-compliant applications | Writing new features, implementing user stories |
| Code Reviewer | Systematic code review | Reviewing PRs, checking standards compliance |
| Tester | BDD-focused testing and Playwright journey tests | Writing unit tests, acceptance tests, journey tests, coverage gaps |
| Accessibility Advisor | WCAG 2.2 AA and GOV.UK Design System compliance | Reviewing or fixing UI templates, forms, and components |
| Orchestrator | Plans and delegates complex tasks to specialist agents | Large features, multi-phase work, multi-agent workflows |
| Repo Setup | Interviews you about your project and generates the full Copilot config | Setting up a new repo, adding Copilot to an existing project |
Writing your own agents
Follow this structure:
---
description: One-line summary of what this agent does
tools: [edit, execute, read, search, web, todos, thinking]
---
# Agent Name
## Role
What this agent is and its expertise.
## Workflow
1. Step one
2. Step two
3. Step three
## Rules
- What the agent must always do
- What the agent must never do
## References
- Links to standards or instruction files this agent follows
Tips:
- Keep the description under 100 characters — it appears in the agent picker
- Prefer tool sets (
edit,execute,read,search,web,agent) over individual tool names — each set bundles all the related built-in tools so the agent has full Agent-mode capability without a long brittle list - Add
todosfor any multi-step agent — it gives the agent (and you) a visible progress checklist - For read-only review agents, omit
editandexecuteto enforce the principle of least privilege - VS Code silently ignores unknown tool names (opens in new tab), so MCP and extension tools can be listed safely
- Reference instruction files rather than restating rules
- Define clear boundaries — what the agent does and does not do
