Tom Ashworth • 5/25/2026

Interlocks: proving the agent read the skill

This article explains how incident.io implements 'interlocks' in their Applied AI codebase to ensure Claude reads the associated skill file before using a tool. It describes two types: skill interlocks, which enforce skill reading by requiring a token, and guidance interlocks, which provide an audit trail of consulted files. The approach helps reduce errors, improve tool usage tracking, and ensure the agent follows intended workflows.

0 comments

#observability #AI Agents #Claude Skills