Skip to content
GitHub Copilot Train-the-Trainer
Slides

Custom Agent Guidance

Bounded roles, explicit tools, human review


Preflight

Before a live exercise, verify current documentation and customer policy for:

  • repository and data boundary;
  • supported surface and profile syntax;
  • permitted tools and MCP access;
  • reviewer, meter, escalation, and stop guard.

Use a non-sensitive starter task. A manual profile review is the fallback.


A role charter

.github/agents/
└── test-writer.md

A profile names one job, allowed inputs and tools, validation, stop conditions, and a manual fallback. It does not grant new authority.


Least privilege

Need Capability Decision
Read target and tests Repository read/search Allow if approved
Add test Edit named test file Allow if approved
Run known test Command execution Conditional
External data MCP or network Remove unless needed
Merge Privileged action Human only

Demo task

Add tests for `validateEmail`.
- `"ada@example.com"` returns `true`.
- `"invalid"` returns `false`.
- `""` returns `false`.
- Change tests only.
- Run the verified focused test command.

Review the diff and test result separately. Stop if production code needs to change.


Pick the smallest extension

Need Use
Repository convention Instructions
Repeatable procedure Skill
Specialized role Custom agent
Approved external capability MCP

Lab

Create and review a test-writer and documentation profile. Use the data-analyst exercise only after its MCP server is approved.