LemonHarness Improves LLM Agent Stability for Long-Horizon Tasks

Large language model agents often struggle with long, iterative tasks due to difficulties in tracking workspace state changes and managing execution. Current systems provide limited visibility into file system modifications and temporary artifact generation, leading to scattered and hard-to-monitor changes. This lack of explicit boundaries can hinder an agent's ability to maintain a coherent operational state. LemonHarness addresses these challenges by creating a unified execution environment. It constrains all state-changing operations, such as file writes and dependency installations, within a clearly defined workspace. This framework also incorporates a reusable rule knowledge base, allowing agents to leverage predefined execution rules and acceptance criteria, and introduces a time-aware mechanism that informs the model about elapsed and remaining budget, enabling better resource allocation. Evaluations on Terminal-Bench 2.0 demonstrated significant improvements. LemonHarness, when paired with GPT-5.3-CodeX, achieved 84.49% accuracy, which further increased to 86.52% with a stronger GPT-5.5 backbone. These results highlight that a structured runtime boundary, accessible rule knowledge, and time-sensitive execution are crucial for more stable and effective long-horizon agent performance.