October 2025 - December 2025.
- Designed and implemented a state-driven Agent Harness framework for long-horizon interactive tasks, addressing the difficulty of maintaining stable task state across extended LLM interactions.
- Abstracted two core components,
AgentRuntimeandTaskSpec, to decouple the general execution loop, tool use, and state updates from task-specific logic. - Supported plugin-style task integration through
TaskSpec, with deterministic task-side state transitions for managing progress and reducing state drift in multi-turn interactions. - Validated the framework with a cooking assistant task, covering intent recognition, RAG-based recipe retrieval, step progression, help QA, progress checks, and completion detection.
- Used conversation summaries to compress long-running task state instead of directly appending full dialogue history, improving context scalability for extended conversations.
- Generated 20 multi-turn interaction trajectories across three scenarios with Gemma4-31B, and compared the Harness agent against a stateless RAG baseline using GPT-5.5 as an LLM-as-a-Judge.
- The Harness agent won 18 out of 20 cases; state management improved from 1.0 to 4.30, multi-turn interaction quality improved from 1.0 to 4.6, while recipe restatement accuracy showed a measured trade-off, decreasing from 4.1 to 3.8.