Why long-horizon agents still drift — and the fix
Agents that run for hours accumulate small errors into big failures. A new memory-anchoring trick keeps them on track.
The drift problem
Long-running agents lose the plot. Each step inherits the small mistakes of the last, and after enough steps the trajectory wanders off.
What changed
Anchoring each decision to a verified history — instead of the agent's own recent output — stops errors from compounding. For builders, this is the difference between a demo and a product.