Show HN: Stop your AI agents from approving their own work
Makerchecker is an open-source Python library that prevents AI agents from approving their own work by requiring human approval for actions. It implements a maker-checker pattern, where a separate person or process must verify changes before they take effect, adding oversight to automated AI workflows.
Background
- "Maker-checker" (or "four-eyes principle") is a long-standing internal-control process where one person creates a request and a second, independent person must approve it before the action is executed. It is standard in banking, accounting, and compliance systems to prevent fraud and errors.
- This GitHub project applies that principle to AI agents: the AI can autonomously "make" an action (e.g., transfer money, delete a record, change a config), but a human must explicitly "check" and approve it before the agent carries it out.
- The project addresses a growing concern as companies give AI agents (autonomous LLM-based software) access to real-world APIs and databases: without a mandatory human approval step, an agent could act on a hallucinated instruction, a prompt-injection attack, or simply an incorrect decision, with no oversight.
- The tool is open-source and uses a middleware layer that intercepts agent actions and blocks them until a human approves or rejects via a simple interface.