Long-horizon terminal tasks differ from short scripted interactions because they require agents to explore, gather evidence, revise plans based on feedback, and turn intermediate work into final deliverables. Standard environment synthesis often shortcuts or completes that substantive work during construction, removing the need for genuine agentic behavior. This paper argues that high-quality long-horizon training data must deliberately preserve and expose those patterns if we want agents that can carry out multi-stage professional work.
Key Findings
- Constructs a focused data synthesis pipeline that emphasizes agentic behavioral patterns rather than only surface tool calls; this yields substantially longer trajectories (average 52.8 turns) that capture exploration, verification, and iterative improvement.
- Builds TerminalHorizon-3K (1,535 task environments and 2,852 trajectories) and uses supervised fine-tuning on Qwen3.5-27B to produce TerminalHorizon-27B, producing double-digit gains across 12 benchmarks including coding, automation, and general reasoning.
- Demonstrates the model's ability to perform multi-stage recursive self-improvement in case studies with 12-hour research budgets, showing continued experimentation beyond initial solutions and retention of effective improvements in final deliverables.
Who it's for and tradeoffs
Great fit if you need training data or methods to teach LLM-based agents sustained, multi-step professional workflows (e.g., coding agents, automation agents, long-horizon research/RSI tasks). The approach is valuable when agent behavior must include exploration, evidence collection, and verification across stages.
Look elsewhere if you only need short, single-step tool-call demonstrations or lightweight instruction-following data; TerminalHorizon prioritizes preserving substantive work and so is more complex to generate and consume than short-turn datasets.
How it works (concise)
The engine has two complementary stages: (1) Challenge-Guided Long-Horizon Synthesis — convert real professional problems into executable tasks and use an independent agent to surface omitted work, then reconstruct environments so that agents must perform that work; (2) Progressive Horizon Expansion — iteratively expand tasks under the same goal to chain additional requirements that depend on earlier results and feedback. The pipeline produces long, connected trajectories intended to elicit agentic patterns and uses them for supervised fine-tuning, yielding measurable benchmark improvements.