Can task state management alone improve long-horizon agent performance?

The paper locates part of the long-horizon failure in where the harness keeps its bookkeeping. Per the abstract, existing harnesses "maintain task execution, task state, and completion assessment within a growing context," which makes the state "difficult to track" and lets "incorrect self-assessments" carry into later decisions. LongHorizon-Harness instead reformulates long-horizon execution as "a task-state management problem," holding the state explicitly outside execution and updating it only with facts independently verified from the environment. The reported results are Qwen 3.7-Plus rising from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench 2.1, and from 2.8% to 8.3% on OSWorld 2.0. The conclusion draws the moral that capability "is determined not only by the underlying model, but also by the harness."

The Manage-Execute-Audit (MEA) loop splits three jobs a single growing context was doing at once. A manager maintains the task state and picks the next subtask, a "fresh-context executor" performs it, and a "read-only auditor" verifies the resulting environment state before the next round. Only "independently verified outcomes" carry across rounds, so an executor's own claim of success does not become state by itself. A lightweight AgentAdapter lets model and harness backends be swapped "without modifying their native agent loops," and the conclusion reports improvements across hybrid GUI–CLI workflows, professional desktop tasks, pure command-line environments, and different model backbones.

Against the nearest notes, this extends Can delegation teach models to manage context more actively?. Both use fresh contexts for subtasks, but that note ends on the caution that a main agent integrates returned summaries it cannot audit, and the MEA loop answers with a read-only auditor that checks environment state instead of trusting the executor's report. That pairing is a reading of the two papers; the excerpt does not cite the delegation work. It contrasts with Can agents discover tools dynamically instead of pre-selecting them?, which keeps a whole-task perspective inside one autonomous stream, whereas MEA holds the whole-task view in the manager and gives each executor a narrow one. It also echoes the factoring in Why do planning and grounding pull against each other in agents?, another case of a harness-level split. And it adds a wrinkle to Does a single benchmark score actually predict agent readiness?: the same model differs by about 29 points on WeaveBench depending on the harness around it, so a reported score describes a model-plus-harness pair.

The excerpt gives only the abstract, one introduction passage and the conclusion. It reports no ablation, so it cannot say whether the gain comes from externalized state, fresh-context execution, or the audit step. It does not name the baseline harness, and it gives no run counts, task counts, or cost of the extra manager and auditor calls. It also does not say how the auditor verifies state where the environment offers little to check. The OSWorld 2.0 result is a large relative gain from a very low base, 8.3% in absolute terms. What the excerpt does support is narrower: harness design is a variable that moves long-horizon scores substantially for at least one model, and it belongs in the report beside the model name.