FACET Synthesizes Executable Terminal Tasks with State Preservation.

Training AI agents for terminal-based tasks requires a large volume of high-quality, executable supervision. However, synthesizing these tasks is challenging because each task involves multiple interconnected components: an instruction, an initialized environment, a reference solution, and an executable verifier. Inconsistencies between these artifacts can lead to unsolvable tasks or incorrect evaluations, and multi-stage synthesis often loses critical information from the original sources. To address these issues, researchers developed FACET (Fine-grained Agentic Construction of Executable Tasks). This framework focuses on preserving the original source intent and maintaining cross-artifact consistency. FACET reconstructs agent skills into rich scenarios, then realizes and repairs the execution environment before generating the final task components. The container state serves as a shared foundation for the instruction, solution, and verifier, ensuring alignment. FACET employs execution-based validation and targeted repair mechanisms to correct specific artifact failures without regenerating valid parts. This approach enables the creation of complex terminal tasks with dense executable checks. The successful trajectories collected from these FACET-generated tasks provide efficient and effective supervision for fine-tuning models, consistently improving performance on benchmarks like Terminal-Bench 2.1 and highlighting the importance of environment-grounded construction for task validity.