This is real, unscripted terminal output from agent_acid, an open-source safety layer for AI agents.
The scenario: an AI agent is told to charge a customer $1,200, but the system only allows charges up to $500 per transaction. So the plan gets split into three $400 charges instead, each one individually legal.
Most AI agent guardrails only check one action at a time, so this would slip right through. agent_acid's shadow execution runs the entire plan in a safe sandbox first, before anything touches a real system. It catches the pattern on the third call and rejects the whole plan.
The real-world result: zero accounts created, zero charges made. Not "created then undone" -- never touched at all.
Compare that to a rollback-only approach, where the account and first two charges would have already happened for real before anything got cleaned up.
Open source, tested, and free to try:
GitHub: github.com/muhammadwaqasai/agent_acid
pip install agent-acid
Curious what people think, especially if you can find a way around it.
Top comments (0)