AI Papers Reader

Personalized digests of latest AI research

View on GitHub

Keeping AI Agents on Task From First Search to Final Deliverable

AI assistants must do more than answer questions. A typical request—“improve this report, add evidence from the web, and hand back a finished slide deck”—can involve dozens of steps across browsers, files, code, and images. Over such long horizons, even large language models forget early constraints, lose intermediate findings, or exhaust their context windows. In a preprint, researchers from Zhejiang University and Ant Group argue that these failures compound, and they propose OneDayAgent, a harness for managing long-horizon tasks.

OneDayAgent works by wrapping a backend LLM in three mechanisms. It decomposes an open-ended request into bounded subtasks; it maintains compact “execution memory” so later subtasks inherit summaries and file handles rather than full interaction logs; and it verifies the final deliverable against the original request, then repairs defects. Each subtask runs in a ReAct loop—reason, call tools, observe results—with access to web search, academic search, Python, file editing, and image tools. When context approaches 90% of the model’s window, earlier rounds are compressed into a technical summary.

That mechanism is visible in a “Language of Flowers” PowerPoint task. The user asked for Wikipedia-based revisions, a comparison table, a Pexels image, deletion of slide 9, and a rewritten conclusion. OneDayAgent split the work into a research subtask and a file-editing subtask; the second failed with a “Bad file descriptor” error. The verifier flagged the missing PowerPoint file and suggested applying the collected research with python-pptx. A repair pass produced the deck, and a second verification confirmed all six required modifications.

On AgentIF-OneDay, a benchmark of 104 daily-life tasks scored across 767 rubric points, OneDayAgent with the GLM-5.2 backend achieved an overall score of 0.821, which the team says leads across every task type, domain, and rubric dimension. Official baselines include Manus at 0.645, ChatGPT-Agent at 0.626, and Minimax-Agent at 0.562; the harness also beat a Codex run using GPT-5.5 medium (0.664) and AutoClaw (0.799). The same harness ran across five backends from three model families, with scores from 0.613 for Qwen3.6-27B to 0.743 for Gemini-3.1-Pro-Preview and 0.821 for GLM-5.2.

Ablation experiments with GLM-5.2 clarify what the harness contributes. With decomposition and verification both disabled, the score was 0.771. Adding either module alone produced scores around 0.804—gains of 3.3 percentage points—and enabling both raised the score to 0.821, a 5.0-percentage-point improvement over the stripped-down version. The combined gain was less than the sum of the isolated gains, suggesting the modules fix some of the same failures.

The extra performance came with a cost. The full harness averaged 53.6 minutes and 51.6 tool calls per task; verification-only averaged 29.7 minutes, close to the 27.6-minute baseline, while decomposition-only averaged 38.1. The authors caution that the results are specific to AgentIF-OneDay and that broader generalization needs further benchmarks. They also note that the current implementation runs without workspace isolation, so malicious web content could, in principle, inject instructions that persist through context compression and lead to unintended shell commands. Those caveats blunt any claim that OneDayAgent “solves” long-horizon agency, but the results suggest a shared harness—rather than a single model—can keep agents anchored to the original goal.