- Published on
The market believes that they can achieve reliable outcomes by releasing a free-run AI agent into a folder (like OpenClaw or Claude Desktop) and giving it access to READ and WRITE a Python script - that performs or checks the math. We tested this. Instead of doing the math, the free-run agent cheated to pass the test! It hallucinated an incorrect output into a file and printed "success", thereby faking the python output. Here is why at Mæstery we put agents on rails and give them tools they can press like a button, masking the python script inside the tool.