Summary
- • LangChain Deep Agents now ships harness profiles for OpenAI Codex, Anthropic Claude/Opus, and Google models out of the box
- • Model-specific prompt and tool tuning yields 10–20 point gains on tau2-bench (multi-turn tool use + instruction following)
- • Harness changes alone moved gpt-5.2-codex from 52.8% to 66.5% on Terminal-Bench 2.0 — no model weight changes required
- • Same model in different harnesses can rank last or Top 5 on leaderboards — harness engineering is now a first-class design variable
Details
Harness profiles ship for OpenAI Codex, Anthropic Claude/Opus, and Google models
Three profiles launch out of the box; the system is designed to be extensible to additional model families. Each profile bundles prompt templates, tool selections, and middleware hooks optimized per model.
Codex profile swaps in apply_patch and shell_command tools per OpenAI Codex Prompting Guide
Aligns the agent harness with the model's training distribution and documented strengths in code manipulation tasks. Tool names and implementations follow OpenAI's own prescriptions.
Opus profile adds tool_result_reflection and tool_usage XML blocks to prompts
Steers Claude models toward grounding reasoning in observed tool outputs rather than memory, addressing a known failure mode in long agentic chains. Changes are prompting-only — no tool modifications.
gpt-5.2-codex jumped from 52.8% to 66.5% on Terminal-Bench 2.0 via harness changes alone
The 13.7-point improvement shifted the model from outside the Top 30 to Top 5 at time of publication — demonstrating that infrastructure configuration can rival model capability differences in measured performance.
10–20 point gains on tau2-bench (multi-turn tool use + instruction following)
Gains of this magnitude are significant; frontier model version upgrades typically yield single-digit benchmark improvements. Measured on a curated subset of more difficult tasks that frontier models haven't yet saturated.
Claude Code harness ranked last among Opus 4.6 submissions on Terminal-Bench 2.0
Illustrates that even state-of-the-art models can be severely harmed by suboptimal harness configuration, and that published benchmark rankings are harness-dependent — not purely a property of the underlying model.
Product Launch = new feature release, New Tech = novel mechanism or tool, Stat = quantitative benchmark result, Insight = analytical conclusion from empirical data
What This Means
Agent developers can no longer treat prompts and tool configurations as generic — this launch provides concrete benchmark evidence that model-specific harness tuning is a primary performance lever, not a finishing step. For practitioners building on LangChain, adopting the appropriate harness profile for their target model is now a baseline expectation. More broadly, benchmark comparisons between models using different harnesses are increasingly difficult to interpret, since the harness itself can account for double-digit performance swings.