Agent mission control

Give a toy agent a task. Step through its tools, budget, approval and evidence. Make it fail safely.

How this experiment works

Observe, inspect, calculate, authorise, seal, act and verify are explicit transitions. The trace shows tool results and policy decisions, not a model’s private reasoning.

What it does not prove

This is a deterministic teaching agent in a fictional world. It does not execute real tools or represent the behaviour of a language model.

Try this challenge

Make the seal fail. Check that summary.txt never changes. Then restore the seal and complete the mission.

Enable JavaScript to run the experiment on your device.