Agent Test Guardrails

Keep AI coding agents from rewriting their own unit tests

Problem

When agents write new features, teams face a dilemma with unit tests: letting the agent update them destroys their value as independent verification, while forbidding it kills velocity as the agent keeps failing old tests. Engineers end up manually diffing which tests were touched, slowly and error-prone.

Solution

A lightweight check between git and CI that flags protected test files and assertions, then diffs test semantics before and after the agent's commit (removed assertions, changed expectations, skipped cases). Violations are blocked and turned into a correction prompt the agent can read, plus a human review list.

Users

Small engineering teams and solo developers using Claude Code, Codex, or Cursor to write production code

MVP

Config and marking of protected test files; assertion-level change detection from commit diffs; blocking with agent-readable correction prompts; a simple web approval panel; GitHub Actions / pre-commit integration

Revenue

Per-repo subscription, free for open source; team tier adds audit history and multi-person approval

Why now

Coding agents are now part of daily workflow, yet there is no accepted norm for whether agents may edit tests; a tool with sensible defaults can become that norm