FACET Targets Silent Failures in Terminal-Agent Task Synthesis

A synthetic training task for a terminal agent has four moving parts: an instruction, an initialized environment, a reference solution, and an executable verifier. Generate them from inconsistent assumptions about the state of the machine and the task is broken, either unsolvable or scored wrong, with nothing in the pipeline to flag it. A new arxiv preprint proposes FACET, or Fine-grained Agentic Construction of Executable Tasks, as the fix.

The framing is direct: "if these artifacts are generated from inconsistent assumptions, the resulting task may be unsolvable or incorrectly evaluated." A second complaint follows, that "multi-stage synthesis can discard the goals, dependencies, state transitions, and procedural constraints encoded in the original sources." Two failure modes, one pipeline.

FACET's move is to make the container itself the shared source of truth. The framework "reconstructs related agent skills into coherent, information-rich scenarios, then realizes and repairs the execution environment before generating the final task artifacts." The instruction, solution, and verifier are then all grounded in that container state. When something breaks, "execution-based validation and targeted repair correct artifact-specific failures without unnecessarily regenerating valid components."

Fine-tuned models "consistently" improve on Terminal-Bench 2.1, the paper reports, though the abstract publishes no per-scale numbers. The claim staked is a design one: source-intent preservation and shared executable-state grounding as first-class principles for building terminal-agent training data.