What Transfers in Agentic Post-Training? Cross-Benchmark Generalization from Long-Horizon Tool-Use Tasks

Lifelong agents must keep improving from interaction without overfitting to the environments they happen to be trained in. We study this plasticity-generality tradeoff empirically: we post-train a 122B mixture-of-experts model (Qwen3.5-122B-A10B) with LoRA-based SFT followed by RL (GSPO) on 363 long-horizon, MCP-based tool-use tasks, and evaluate transfer across five benchmarks spanning tool orchestration and software engineering. Training only on office-style tool-use tasks improves not just in-distribution holdouts (+17.3pp) and Toolathlon (+9.6pp), but also SWE-Bench Pro (+6.3pp) and Terminal Bench 2 (+5.2pp), domains never seen in training. Trajectory analysis suggests the gains come from reusable goal-directed reasoning: forming precisely grounded subgoals, building goal-relevant working state, preserving parent-goal constraints during local work, and verifying completion against the original objective. We also report negative results for continual training infrastructure: sparse rewards provide signal on too few tasks to train on (dense partial-credit rewards raise the usable task set by up to 81pp), and live MCP services fail in ways that silently corrupt training signal. We distill these into concrete design recommendations for lifelong agent training environments.