OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents
Jingsheng Zheng, Xinyuan Fang, Jintian Zhang, Zhengke Gui, Huajun Chen, Ningyu Zhang
cs.CL, cs.AI, cs.HC, cs.LG, cs.MA
2026-08-05
A decompose, memory, verify/repair harness hits 0.821 with GLM-5.2 on 104 AgentIF-OneDay tasks (beating 0.799) and lifts every one of five backends tested.
Open-ended everyday tasks (booking an itinerary, assembling a report with attachments) are hard for agents because they are long-horizon, cross-environment, and multimodal. The model has to hold goals and constraints across many steps while switching heterogeneous tools and attachments, and three failure modes pile up: goal drift (a formatting requirement from early on gets dropped), state loss (search evidence found in one environment does not carry into the local-file environment), and context overflow (the context blows up before the deliverable is done).
Prior work picks these off one at a time with reasoning scaffolds, feedback-based revision, or memory management. The trouble is that the failures amplify each other, so fixing one in isolation is not enough. Whether a single harness can manage them jointly and still work across backends had not been tested head-on.
OneDayAgent is a harness wrapping the LLM backend; the backend reasons, the harness runs the process. Three mechanisms do the work.
Task decomposition. The harness turns the original request into an ordered list of bounded subtasks, each an executable unit that can call tools, produce files, and submit a compact answer. The backend faces one short local objective at a time, while the harness keeps the original request as the global intent. The subtask boundary also saves context: later subtasks inherit task-level state but not the full low-level ReAct trace.
Execution memory. Three sub-mechanisms absorb context pressure: search results become structured snippets, long pages become bounded summaries, file reads become modality-aware previews; subtasks exchange compact checkpoints (submitted answers and file handles) so evidence is reused across environments without the full trace; once the accumulated trajectory crosses a fraction of the backend window, earlier rounds are compressed into a technical summary while the system prompt, original task, and recent rounds stay verbatim, with deterministic emergency pruning near the hard limit.
Global verification and repair. After synthesis, the harness checks the candidate deliverable against the original request, the subtask answers, and the declared attachments. It is artifact-level and task-global, not just the final text. On a defect it runs targeted, ReAct-style repair from the verifier's defect description rather than restarting every subtask, then re-verifies. Completing every subtask does not guarantee the deliverable meets the original request; long-horizon runs can still lose early constraints or skip implicit requirements.
Evaluation runs on AgentIF-OneDay, 104 tasks and 767 instance-level scoring points across work, study, and life. With GLM-5.2, OneDayAgent hits an overall 0.821, a new state of the art.
| Method | Overall |
| AutoClaw (best baseline) | 0.799 |
| Manus | 0.645 |
| ChatGPT-Agent | 0.626 |
| OneDayAgent (GLM-5.2) | 0.821 |
The same harness across five backends scores 0.613 to 0.821: GLM-5.2 highest, Gemini-3.1-Pro 0.743, Qwen3.5-397B 0.708, Qwen3.5-9B 0.624, Qwen3.6-27B 0.613. The ordering is not monotonic; Qwen3.6-27B does not beat the smaller Qwen3.5-9B, so the win is more about backend-harness fit than scale.
Ablations run on GLM-5.2. Execution memory cannot be turned off; doing so triggers context overflow and tasks do not finish. Decomposition alone reaches 0.804, verification alone 0.804, both together 0.821, all above the 0.771 base. The combined gain is smaller than the sum of the two isolated gains, so the modules recover partly overlapping failures. On cost, verification wins clearly: it matches decomposition's score for 2.2 extra minutes, while decomposition adds 10.6 minutes and roughly 60% more tool calls. Of 104 tasks, 95 pass verification on the first attempt and 9 enter repair; 6 of those are recovered.
For agent builders the takeaway is engineering: on long-horizon everyday tasks the win is in the harness, not the model. One unchanged decompose-memory-verify/repair shell lifts every backend tested and beats single-trajectory ReAct. The cheapest single ingredient is the final verify/repair pass, low cost for decomposition-level gain. Swapping backends is not silent; different models run different execution styles under the same workflow, so selection needs real measurement.
The authors mark several bounds. Findings hold on AgentIF-OneDay and need more benchmarks to generalize. Cross-backend transfer is not silent (scores span 0.613-0.821) and full is not always best; verification outscores full on 17 tasks. The link between context management and stable quality is not causally isolated. On safety, the implementation runs on the host with no workspace isolation, executes arbitrary shell commands with no allowlist, and a prompt injection that survives compression can leak into later subtasks. The judge was also swapped: the official Gemini-3-Pro-Preview was unavailable in June 2026, replaced by the stricter Gemini-3.1-Pro-Preview, which the authors say makes their numbers conservative.
One more thing for the reader to weigh: the paper is from zjunlp at Zhejiang University, the headline backend GLM-5.2 is from Zhipu, and the GLM-5.2 run is the highest-cost profile (53.6 minutes, 51.6 tool calls, 585.7 KB per task) producing the 0.821. Read the 2.2-point lead over AutoClaw's 0.799 with that in mind.