LangChain improves coding agent performance on Terminal Bench 2.0 by refining harness engineering

LangChain's coding agent, using the fixed model gpt-5.2-codex, improved its score on Terminal Bench 2.0 from 52.8% (outside Top 30) to 66.5% (Top 5) solely by adjusting the agent's harness. This harness optimization focused on three knobs: System Prompt, Tools, and Middleware, utilizing automated trace analysis from LangSmith to identify and correct failure modes across 89 tasks.