WorkBorne introduces a synthesis framework that treats artifact transitions as the primitive for modeling agentic tasks. The pipeline first builds coherent and diverse artifact corpora via iterative web retrieval; each corpus contains real files, configurations, or data collections. It then discovers non‑trivial input‑output artifact assemblies and infers the semantic relations linking them, such as "edit", "deploy", or "migrate". Every transition records the performer, the initial artifact, the intended outcome, and their relationship, and is instantiated into a grounded task runnable in a terminal environment.
The authors construct 1,061 artifact corpora and 7,882 task environments, yielding 15,485 high‑quality trajectories, with 53.6% exceeding 30 interaction turns. Evaluation with Qwen3.5‑27B and Qwen3.5‑35B‑A3B raises Terminal‑Bench 2.1 success from 43.1% to 61.8% and lifts the ClawEval General score from 75.0 to 80.2, outperforming all prior synthesis baselines. Further analysis shows transition depth strongly correlates with execution depth, while corpus breadth and transition depth provide complementary scaling benefits.
Review