What Transfers, and Why: A Behavioral Account of Cross-Benchmark Generalization in Long-Horizon Agents

Agentic post-training is usually judged by whether task scores go up. For reinforcement learning (RL) in a self-built environment this is a weak signal: a policy can raise its score by exploiting environment-specific regularities (tool schemas, grader parsing, task templates) rather than by acquiring transferable skill, and an in-distribution holdout shares exactly those regularities. We argue that the discriminating question is behavioral, namely how a trained agent acts, and that cross-benchmark transfer is the right place to look for it. We post-train an open-weight Mixture-of-Experts model (Qwen3.5-122B-A10B) on a long-horizon RL environment of 363 Model Context Protocol (MCP) tasks across 27 categories, using a two-stage SFT-then-RL pipeline with dense, per-criterion rewards. The trained model improves over the base on five independently constructed external benchmarks it never trained on: Toolathlon (+9.6 pp), τ²-Bench (+5.3 pp), BFCL-V4 (+3.5 pp), SWE-Bench Pro (+6.3 pp), and Terminal-Bench 2.0 (+5.2 pp). The improvement is directionally consistent (sign test p = 0.031) and holds for software-engineering benchmarks despite no coding-specific curriculum. We then ask why. Paired trajectory analysis isolates four recurring behavioral changes (more careful goal formation, building goal-relevant working state, keeping parent goals stable through local repairs, and verifying completion) that recur in the same form across spreadsheets, terminals, and code, and we quantify them on software trajectories. We separate genuine goal-directed reasoning from a shallower "learn to attempt and emit artifacts" propensity, and we characterize where transfer fails: preserving existing behavior in mature codebases. We report binomial confidence intervals, treat only the large and directionally consistent base-to-trained gains as robust, and discuss the single-run limitation of our RL result.