OpenHarness drops code-first agent orchestration SDK

OpenHarness shifts the industry focus from raw model scores to the actual reliability of the system surrounding the LLM.

• –Framework treats the "harness" (the system around the model) as a primary variable in performance, enabling teams to benchmark how orchestration impacts final outcomes.
• –Deterministic playback mode allows for reproducible testing of non-deterministic LLM logic, ensuring that evaluation runs are consistent in standard CI/CD pipelines.
• –Native Model Context Protocol (MCP) support facilitates standardized tool-use benchmarking across different model providers and development environments.
• –Two-phase history compaction (pruning and summarization) prevents performance degradation in long-running sessions, addressing a critical metric for production-ready agents.
• –Lightweight TypeScript SDK approach makes it significantly easier to integrate into existing production codebases compared to "heavyweight" agent frameworks.