Jingsheng Zheng, Xinyuan Fang, Jintian Zhang, Zhengke Gui, Huajun Chen, Ningyu Zhang
6 min
As autonomous agents move from simple question-answering to complex, real-world workflows, they face significant challenges in maintaining goal alignment and state consistency over long periods. The authors investigate whether a unified 'harness'—a structural framework for managing agent execution—can solve the compounding failure modes of goal drift, context overflow, and state loss in open-ended, multi-environment tasks.
OneDayAgent introduces a structured execution loop that moves beyond standard ReAct-style prompting. The harness implements three core capabilities:
These modules operate over a unified tool interface that supports web search, file manipulation, code execution, and multimodal processing.
The authors evaluated OneDayAgent on the AgentIF-OneDay benchmark, which consists of 104 complex, multi-step tasks. Using the GLM-5.2 backend, OneDayAgent achieved an overall score of 0.821, setting a new state of the art. The study demonstrates that the harness is model-agnostic, maintaining stable performance across five different LLMs from three distinct model families without requiring backend-specific tuning. Ablation studies confirm that both the decomposition and verification modules contribute significantly to the overall success rate, justifying the additional computational overhead.
This research addresses a critical bottleneck in agentic AI: the transition from 'chatting' to 'doing.' By providing a modular harness that manages the lifecycle of a task, OneDayAgent offers a blueprint for building agents that can reliably handle multi-hour, cross-environment workflows—such as researching a topic, editing local files, and generating a final report—without losing track of user constraints or intermediate findings.
LLM agents are increasingly applied to open-ended everyday requests that span work, study, and life. These tasks are long-horizon, cross-environment, and multimodal, forcing the agent to preserve goals and constraints across many steps while navigating heterogeneous tools and attachments. While prior work has addressed individual failure modes such as goals drift, states loss, and context overflow, whether a single harness can manage them jointly and remain effective across backends has received less study. We present OneDayAgent, a long-horizon harness for autonomous agents. OneDayAgent turns an open-ended request into a managed execution process that decomposes tasks into bounded subtasks, maintains execution memory under context pressure, and verifies and repairs the final deliverable. We evaluate OneDayAgent on AgentIF-OneDay across 104 tasks. With the GLM-5.2 backend, OneDayAgent sets a new state of the art with an overall score of 0.821. The same harness runs across five backend LLMs from three model families, indicating the harness generalizes across backends without tuning, even as different models induce distinct execution styles under the same workflow.
Alex: That's a useful analogy. And then the third part kicks in at the end: a "global verifier." Once the full task is complete, the system compares the finished work against the original instructions to check whether everything was actually done correctly.
Sam: And if something's wrong—does it start over from scratch?
Alex: No, and that's an important design choice. It uses what they call "targeted repair." Rather than throwing out all the work and beginning again, it identifies specifically which part failed and fixes only that piece. It's the difference between rewriting an entire essay because one paragraph is weak, versus just rewriting that paragraph.
Sam: That's a meaningful efficiency gain. Did this approach actually produce better results in practice?
Alex: It did. When tested on a purpose-built evaluation called the "AgentIF-OneDay" benchmark, the harness performed meaningfully better than baseline agents that lacked this structure. And one finding worth highlighting: the improvement held across five different AI model families, without any special tuning for each one. The harness itself is doing the heavy lifting—it's not dependent on one particular AI being especially capable.
Sam: So the system works regardless of which AI engine is underneath it. The management layer is what matters, not the specific model.
Alex: That's the central claim. And it has a practical implication: as AI models improve over time, you could swap in a better one and the harness would still function. The infrastructure and the intelligence are kept separate.
Sam: The paper is also candid about what doesn't work yet, which I appreciate. There's a security concern they flag quite openly.
Alex: Yes. In its current form, the agent runs directly on the host computer without any kind of protective barrier around it. Because it can execute commands and browse the web, it could potentially be manipulated by malicious content it encounters—a hidden instruction on a webpage, for instance, could cause it to take actions it shouldn't. The researchers describe this as a priority for future work.
Sam: So without that quarantine layer—what you'd call a "sandbox"—the agent essentially has open access to the machine it's running on.
Alex: Correct. They were candid that this was an infrastructure constraint rather than an oversight. They couldn't get a stable isolated environment working in time for these experiments, so they focused on validating the harness logic first. It's a reasonable trade-off for a research paper, but it does mean the system isn't ready for deployment in sensitive environments.
Sam: Are there other limitations they flag?
Alex: There's also a subtler issue with how performance is measured. Different AI models, when used as judges to evaluate the agent's output, apply different standards. A stricter judge naturally gives lower scores. When the researchers swapped to a more demanding evaluation model, their scores dropped noticeably—which actually suggests their primary results are on the conservative side rather than inflated.
Sam: So the headline number is more likely a floor than a ceiling. That gives the findings more credibility, not less.
Alex: That's a reasonable way to read it. The paper is careful not to overstate what's been shown. What they've demonstrated is that structuring the process around an AI—breaking work into steps, maintaining a running memory, and verifying the output—produces more reliable results than simply prompting a capable model and hoping for the best.
Sam: It's a useful reframe. We tend to talk about AI progress in terms of how smart the model is. But this suggests the system around the model matters just as much.
Alex: And that's likely where a lot of practical progress will come from—not just more powerful models, but better structures for putting them to work. Thanks for listening to ResearchPod.