Show HN: Diff two versions of an AI agent and catch silent permission changes
The article introduces Drift, a tool that lets users compare two versions of an AI agent to detect silent changes in permissions or behavior, helping catch unintended modifications before deployment.
Background
AI agents (also called LLM agents) are programs that give a large language model (like GPT-4 or Claude) the ability to call external tools, read files, or access the internet. Because agents are typically defined in code or configuration files, even a small change — such as an altered system prompt or a new function — can silently grant the AI permission to do something a previous version couldn't (e.g., write to disk, send emails, or modify a database). This project, called Drift, is a diff tool specifically for agent configurations: it compares two versions side by side and flags any changes that expand the agent's permissions or capabilities. It matters because teams building AI agents often lack good code review practices for prompt or tool definitions, and a seemingly harmless update could introduce security or safety risks that go unnoticed in a standard code review.