OneDayAgent Posts 0.821 on 104-Task Everyday-Agent Benchmark
TL;DR
- OneDayAgent turns open-ended requests into a managed process that decomposes tasks into bounded subtasks, maintains execution memory, and verifies and repairs deliverables.
- On the authors' AgentIF-OneDay suite of 104 tasks, the GLM-5.2 backend scores 0.821, which the paper calls a new state of the art.
- The same harness runs across five backend LLMs from three model families without per-model tuning, though the paper notes distinct execution styles emerge.
A new paper on arxiv this week points at something more interesting than another leaderboard number, even though it also has one of those. A research group reportedly built a single agent harness that runs across five different LLM backends without per-model tuning, and posted a top score on their own long-horizon benchmark using one of them.
The setup is called OneDayAgent, and the paper on arxiv frames the target as the boring-sounding but genuinely hard version of agent work: open-ended requests that span work, study, and life, over many steps, across heterogeneous tools, with mixed modalities. The failure modes the authors call out are the ones anyone who has tried to run a real agent has hit, goals drift, states loss, context overflow. Their harness turns a request into a managed process that decomposes tasks into bounded subtasks, maintains execution memory under context pressure, and verifies and repairs the final deliverable.
The headline number is 0.821 on AgentIF-OneDay, a 104-task suite the authors built alongside the harness. That specific score belongs to the GLM-5.2 backend, which the paper calls a new state of the art on the benchmark. The claim that matters more, if it holds up, is the second one: the same harness ran across five backend LLMs from three model families with no per-model tuning, even as different models induced distinct execution styles under the same workflow.
The honest caveat is that this is the authors' harness scoring on the authors' own benchmark, and the arxiv listing flags it as ongoing work, so independent replication is still to come. The abstract does not name the other four backends or the three families, or report how they scored relative to GLM-5.2, so the backend-agnostic story is directional rather than settled. If it does hold, the interesting shift is that agent moat moves from picking the right model toward designing the harness around it, which is a very different competitive game for anyone shipping agentic products.
Originally reported by paper
Read the original article →Original headline: OneDayAgent Introduces 104-Task Full-Day Real-World Benchmark, Claims SOTA 0.821 Across 5 LLMs Without Per-Model Tuning