Reinforcement learning now trains language‑model agents that act for dozens of steps in live environments, yielding large gains that are often read as better decision‑making. Because an agent writes its own inputs in a closed loop, each observation depends on its earlier actions, so an SFT checkpoint and an RL checkpoint are evaluated from different states even on the same task. Endpoint success therefore mixes two factors: where the agent arrives and what it does after arriving. Restricting comparison to states both policies reach does not separate them; this restriction selects on an outcome and, in our data, flips the sign of the effect. We introduce a checkpoint‑handoff protocol: clone the state reached by one released checkpoint and hand it to another checkpoint without any retraining. By crossing a “reacher” role and a “solver” role across SFT and RL, the endpoint gain is split into REACH and SOLVE. REACH measures how often a policy arrives at a state that the environment confirms is a fixed number of actions from success; SOLVE measures how often it finishes from an identical cloned state. Experiments on two benchmarks and two independently released pipelines show a positive interaction in all five conditions. An RL history is more valuable to an RL solver than to an SFT solver. On ALFWorld, RL improves both terms, and the SFT solver never succeeds where the RL solver fails. Independent REACH and SOLVE gaps predict the aggregate interaction. Handoff only requires that one checkpoint’s history can be replayed under another, allowing long‑horizon evaluation to report arrival and completion alongside endpoint success.
Review