What Makes Interaction Trajectories Effective for Training Terminal Agents?

arXiv:2606.03461v1 Announce Type: new Abstract: It is frequently presumed that more capable code agents serve as better teachers for post-training; however, this notion is often confounded by variables such as task complexity, harness architecture, and the student model’s inherent capacity. To clarify this pedagogical relationship, we employ Terminal-Lego, a scalable framework designed to convert real-world, multi-domain challenges into agentic tasks that are verified by their environments. Our findings challenge the conventional wisdom that raw performance correlates with teaching quality. Although Claude Opus 4.6 secures higher rankings on Terminal-Bench 2.0, student models fine-tuned using trajectories from DeepSeek-V3.2—a lower-performing agent—demonstrate markedly superior generalization capabilities. We explain this "pedagogical paradox" through the concept of Environment-Grounded Supervision (EGS). This approach emphasizes trajectories that make inspect-act-verify behaviors visible within the harness, enabling students to absorb durable problem-solving strategies rather than brittle action patterns. Our scaling analysis highlights remarkable data efficiency: utilizing just 15.3k Terminal-Lego trajectories, Qwen3-32B reaches a 24.3% score on Terminal-Bench 2.0, matching previous state-of-the-art results achieved with more than 30 times the data. These insights indicate that the next stage of agent post-training extends beyond simple outcome alignment, pivoting toward "Harness Engineering." In this paradigm, the deliberate structuring of environment-grounded interactions becomes the key driver for creating reproducible and generalizable agentic intelligence.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC