Harness-only: Terminal Bench 52.8 → 66.5 without touching the model
Feb 17 Trivedy (LangChain): deepagents-cli hill-climbs via traces, PreCompletionChecklist self-verify, loop detection, and reasoning sandwich — same gpt-5.2-codex throughout.
Vivek Trivedy’s Feb 17 LangChain post is a clean receipt that the harness is the product: deepagents-cli moved from 52.8 → 66.5 on Terminal Bench 2.0 (+13.7, into the Top 5) while keeping the model fixed at gpt-5.2-codex. No weights swapped — system prompt, tools, and middleware (hooks around model and tool calls) did the climbing.
The improvement loop is itself agentic. A Trace Analyzer skill pulls LangSmith experiment traces, spawns parallel error-analysis agents, synthesizes failure modes, and proposes harness patches — boosting on yesterday’s mistakes rather than re-reading vibes. The dominant failure they saw: the agent writes a plausible solution, re-reads its own code, declares victory, and stops. The fix is structural, not polite. Prompting forces a plan → build-with-tests → verify-against-spec → fix cycle; a PreCompletionChecklistMiddleware intercepts exit and re-injects a verification pass against the task spec (Ralph-Wiggum-shaped: the harness refuses “done” until the checklist runs). LocalContextMiddleware injects cwd maps and available tooling so the agent wastes fewer turns rediscovering the environment. LoopDetectionMiddleware counts per-file edits and nudges a rethink after N edits to the same file — a short-horizon patch for today’s doom loops that should dissolve as models improve. Reasoning spend is budgeted as an xhigh–high–xhigh sandwich (plan and verify expensive; mid-run cheaper); all-xhigh scored worse (53.9%) on timeouts than a tuned split (66.5%).
They are explicit that Claude Opus 4.6 on an earlier harness landed 59.6% — competitive, but worse than Codex because they did not re-run the same improvement loop for that model. Harnesses are model-specific surfaces; verification and context prep generalize, but the hill-climb must be re-run.
Why it matters for code review
Overnight and always-on factories compound only when “done” is a harness decision, not a narrator mood. This note’s measured claim is the week’s thesis in numbers: modest verified momentum — traces in, checklist gates, loop breaks — beats peak-session mastery of a bare chat window on the same model. FOMO when nothing ran overnight is often missing middleware, not missing genius.