Google open-sources Harness²: recursive agent harnessing lifts GLM-5 and Gemini 3.5 Flash by up to 11.8 points
dr_cintas · x · 2026-10-18
Google Research open-sourced Harness², a plug-and-play framework for recursive agent harnessing in an open world, with model weights kept fixed.
Core idea: a harness is the system around a model that makes it an agent (prompts, tools, memory, skills, orchestration, etc.), organized into 8 components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins.
Two levels of recursion:
- Task-level recursion refines the execution harness per task via a contrastive Propose–Probe–Compose step: propose edits, probe their effects on the same task, and compose a refined harness from behavioral contrasts — without ground-truth feedback.
- Harnessing-level recursion maintains a persistent improvement harness with adopted harnesses and evidence-backed lessons from each task to guide future refinement.
Results: across three professional benchmarks and four model–harness configurations, a single Harness² step improves GLM-5 and Gemini 3.5 Flash over their base harnesses — up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB's partial pass rate; larger recursion budgets yield further gains. Supports parallel/sequential recursion plus OpenCode and Codex executors.
More from coding & agent
- AI-driven compiler optimization spree: JSON.parse up to 2.9x faster, shift now linear — cramforce · 2026-10-18
- Papers with Code adds Auto-Research category, 166 papers already tracked — NielsRogge · 2026-10-18
- Use Claude to recursively optimize deps for native tsc — original types are on GitHub — cramforce · 2026-10-18
- Claude Code shows fake summarized thinking as assistant messages — ~37% of transcript — Sauers_ · 2026-10-18
- Dev builds a demon-summoning TUI for managing AI agents — voidwalker_com · 2026-10-18
- Fix wandering coding agents with a DOMAIN_INVARIANT.md file of domain rules — 0xkarasy · 2026-10-18