watch it work
The agent wanders off its goal one small step at a time. A monitor watching only its recent steps sees each one as fine — it never fires. A fixed reference watches the distance from the goal grow, catches it, and returns the agent to ground (the green path). That detect-and-return loop is the harness — and there’s a proof the catching is something self-watching can’t do.
Toggle the harness. On, the reference catches the drift and returns the agent to ground; off, it still detects but nothing acts — the agent spirals past the threshold. Self-watch stays “on track” either way.
Notice the left gauge never moves — out or back: step to step, the agent always looks fine to itself, which is why anything watching only its recent history is provably blind to slow drift. The fixed reference on the right is the only thing that catches it, and once it does, the harness returns the agent to its goal (the green path). laserbrain is that reference — one line for any agent to attach. Whether the return keeps the final answer as good is the honest open question; the research has the studies, nulls included.