OneDayAgent Sets AgentIF-OneDay Record by Treating the Harness as the Product
OneDayAgent scores 0.821 on AgentIF-OneDay with GLM-5.2 and generalizes across five backends — a harness-first datapoint for long-horizon agent research.
Long-horizon agents fail for predictable reasons: goals drift, state gets lost, and context windows overflow. A new preprint argues the fix is not always a bigger model — it is a harness that manages decomposition, memory, and verification as one system.
Researchers introduced OneDayAgent, a long-horizon harness for autonomous agents, in arXiv:2608.05013. The system turns open-ended requests into a managed execution process: bounded subtasks, execution memory under context pressure, and final deliverable repair.
Benchmark results
On AgentIF-OneDay across 104 tasks, OneDayAgent with a GLM-5.2 backend scored 0.821 overall — a new state of the art on that benchmark at publication. The same harness ran across five backend LLMs from three model families without tuning, suggesting the workflow generalizes even when models induce different execution styles.
Why harness design matters now
Agent benchmarks increasingly mix web, file, computation, and multimodal tools in single sessions. Prior work often addressed individual failure modes — goal drift, lost state, context overflow — in isolation. OneDayAgent's claim is that a single orchestration layer can address them jointly while remaining backend-agnostic.
The paper evaluates decomposition quality, memory retention under pressure, and verification/repair loops as co-equal components rather than optional add-ons.
Takeaways for practitioners
The authors highlight two implications: first, long-horizon reliability may depend as much on harness engineering as on raw model capability; second, benchmark scores should be reported at the model-plus-harness configuration level, not model alone.
That framing aligns with a broader August 2026 trend — papers on LongHorizon-Harness, EASy, and Argus all stress runtime orchestration as the binding constraint on real-world agent tasks.
Limits and open questions
The preprint focuses on AgentIF-OneDay; transfer to safety-critical or adversarial environments remains an open question. Verification loops help, but they do not replace domain-specific guardrails for high-stakes actions.
Still, OneDayAgent offers a concrete datapoint: on long everyday tasks, the harness can matter as much as the backbone model — and the gap shows up directly in published scores.
Sources
- arXiv — OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents (Aug. 2026)