Stanford open-sources Meta-Harness: evolving agent scaffolds pushes Opus 4.6 to 76.4% on Terminal-Bench 2.0
victor_explore · x · 2026-11-10
Stanford's IRIS lab open-sourced Meta-Harness (paper 2603.28052, 1.8k GitHub stars): a framework that automatically searches over task-specific harnesses — the code around a fixed model that decides what to store, retrieve, and show — letting an agent rewrite its own scaffold instead of retraining the model.
Key points:
- Motivation: coding agents waste their first 2-5 turns on exploration like ls and which python3; a better harness eliminates this
- The evolved harness lifted Claude Opus 4.6 to 76.4% on Terminal-Bench 2.0 (89 tasks × 5 trials; Easy 100 / Medium 81.1 / Hard 64.7), built on Terminus-KiRA and Harbor
- Repo ships a reusable framework plus two reference experiments (memory-system search for text classification), MIT licensed
It's a notable datapoint for the "harness engineering" route: keep the model fixed, gain big from automated search over surrounding code.
More from coding & agent
- Rerun: open-source AI coworkers that spin up a digital employee team from one prompt — aigclink · 2026-11-10
- Kylon goes cross-platform with 150+ agent use cases across 15 industries — testingcatalog · 2026-11-10
- Codex rolls out global dictation using ChatGPT's fast transcription path — gabrielchua · 2026-11-10
- Zero-cost uncensored deep research stack: Ollama + SearXNG + dual-model fact-checking in Hermes — damnGruz · 2026-11-10
- Open-sourcing PAIR: building trustworthy evals on your own data, fair across vendors — kalyan_kpl · 2026-11-10
- Bloomberg's AI coding interview: 30 minutes on a huge Django codebase, still rejected — dotey · 2026-11-10