OneDayAgent
A long-horizon agent harness that decomposes an open-ended, multi-hour request into bounded subtasks, verifies each deliverable against the original goal, and compresses execution state into reusable memory.
OneDayAgent turns an open-ended everyday request, such as researching a topic, building a slide deck and delivering a report, into a managed execution process instead of a single long-running agent call. It decomposes an overloaded request into bounded subtasks, each with a shorter local objective that still preserves the original global intent, which keeps the context load on each step small and lets progress be checked as it happens rather than only at the end.
After each subtask runs, an LLM-based verifier checks whether the deliverable actually satisfies the original request, and repairs it if defects are found, while execution state is compressed into reusable memory to survive the sustained context pressure of a task that stretches from minutes to hours across environments. Evaluated on the AgentIF-OneDay benchmark of 104 tasks across five backend LLMs from three model families, it reports an overall score of 0.821, ahead of the compared baselines, and its trajectory dataset is published on Hugging Face.