Scaling Agentic Data for Long-Horizon Terminal Intelligence
TerminalHorizon is a data engine for training terminal agents on longer, staged professional tasks.
The paper says short command traces miss how agents gather evidence, revise plans, verify work, and turn partial results into deliverables. TerminalHorizon starts from real-world work, checks what key work was skipped in synthetic environments, and rebuilds tasks so the agent still has to explore and implement. The authors used it to create 1.5K task environments and fine-tuned Qwen3.5-27B into TerminalHorizon-27B. They report gains across 12 benchmarks and 12-hour case studies where the model kept experimenting past an initial solution. HF Daily Papers' note
The paper says short command traces miss how agents gather evidence, revise plans, verify work, and turn partial results into deliverables. TerminalHorizon starts from real-world work, checks what key work was skipped in synthetic environments, and rebuilds tasks so the agent still has to explore and implement. The authors used it to create 1.5K task environments and fine-tuned Qwen3.5-27B into TerminalHorizon-27B. They report gains across 12 benchmarks and 12-hour case studies where the model kept experimenting past an initial solution. HF Daily Papers' note
score 5