Chapter 7: Walkthrough: Copilot Scaffolds the Safe Parts¶

Chapter 7 hero image

Demo videos¶

This is Chapter 5's idea made concrete. In a few minutes, GitHub Copilot can scaffold a new tool for the Hawker agent, add a guardrail, and write the tests. Follow along step by step.

Step by step¶

1. Scaffold a typed tool. Prompt Copilot for a function with a typed signature, input validation, and a Pydantic return type:

Add find_stalls(location: str, cuisine: str | None = None) -> list[Stall]
that calls the Maps API tool. Use the Stall pydantic model from models.py.
Validate location is non-empty. Type the return. Raise InvalidInput on bad input.

2. Add a guardrail (policy as configuration). Wrap the tool with the governance decorator so the policy is loaded from YAML, not hardcoded:

from merlions.governance import govern, load_policy

@govern(load_policy("hawker"))
def find_stalls(location: str, cuisine: str | None = None) -> list[Stall]:
    ...

The hawker policy sets the allowlist, blocked patterns, a call limit, and require_citation. Your security team can change it without a code change or deployment.

3. Generate the unhappy-path tests first. Ask Copilot for empty input, malicious input (an API key in an argument), tool failure with retry, and a happy path. Run them:

pytest tests/test_find_stalls.py -v   # all green

The lesson¶

Copilot didn't write the agent. It wrote the scaffolding so you could focus on agent behaviour.

Notice the order: unhappy paths first. That is what catches the bugs that ship to production, including the fail-closed assertion that an ambiguous policy match denies the call.

Key terms¶

@govern decorator: wraps a tool so every call is checked against the policy and audit-logged.
Fail-closed test: a test that proves the system denies when uncertain.
Mock: a stand-in for an external API so tests are fast and deterministic.