Gibson Runtime

We built Gibson because we kept finding the same thing on every assessment. The agent was fine. The model was fine. The thing that was missing was a line the agent could not cross, and a record of every step it took up to that line.

So there are two ideas in it, and only two.

A grant, not a prompt. A named person gives an agent the rights it may use: read, write, and execute. The rights are bounded by what that person holds. The person sets them once, up front. There is no runtime approval pop-up, because a pop-up is a prompt with a button on it.

Enforced, not advisory. A control that inspects the output and scores it afterwards is advice. A control that makes the disallowed thing unrepresentable is a guarantee. Gibson only claims the second kind.

Everything else is plumbing for those two ideas. Every agent gets an identity. Untrusted work gets a sandbox. Every session gets a budget. Every run gets a journal you can replay for the person who asks what happened.

Our sibling company, zeroroot.ai, builds and sells Gibson. We teach on it because we wrote it. You do not need it to take a course, and you do not need a course to run it.