Definition
Tools that watch what an AI agent actually does — its reasoning, tool calls, and the real-world effects of its actions — to spot when it starts acting outside its purpose. Leading AI labs now publish layered monitoring stacks, and the UK's AI Security Institute documented a real incident where an agent's reasoning bypass led to actions beyond its limits.
Why it matters
Boards need confidence that an agent which has gone off-script will be caught and stopped; this is the difference between 'the model answered oddly' and 'the agent took an unauthorized action that changed a system.'