AI Pulse · Under the Radar

A new open source project called decionis/agent-safe-pipeline popped up five days ago and already has over 500 stars, picking up more than 100 a day. That kind of speed usually means one thing. People building AI agents are all bumping into the same problem at the same time, and this project is one of the clearer answers to it.

The idea is simple once you see it. An AI agent doesn’t just go do things on its own. It proposes an action first, and that proposal gets locked so nobody can quietly change it later. A separate policy check reviews it and hands back one of three answers: allow it, escalate it to a person, or block it outright. If a human needs to sign off, they do, and the agent gets a single-use permission slip to carry out that one action, nothing more. Then an executor actually runs it. Propose, verify, approve, grant, execute. No step gets skipped.

Here’s the part worth noticing. This is the third separate team in under two weeks to build almost this exact structure from scratch, with no obvious coordination between them. When three unrelated groups land on the same design independently, that’s usually a sign the pattern is right, not a coincidence. You can find the project on GitHub if you want to see the actual code.

For any owner letting AI agents touch real systems, invoices, customer data, or scheduling, this is the shape you want underneath, even if you never look at the code. An agent that can suggest an action but not silently execute it is the difference between a helpful assistant and a liability. This is exactly the kind of thing we build into an AI command centre, where every agent action gets logged, checked, and approved before it touches anything real.