Google's Harness²: Recursive Self-Evolving Agent Execution Framework, Up to +11.8 Points
google · hf · 2026-10-14
Northwestern and Google Cloud AI Research publish Harness²: Recursive Agent Harnessing for an Open World, arguing that an LLM agent's capability depends not only on the underlying model but also on its execution harness, and proposing recursive self-improvement (RSI) to automatically evolve the harness:
- Core challenges: open-world tasks are unseen at design time with no ground-truth verifier to guide online modification; conventional RSI is highly task-specialized and hard to generalize
- Two-level recursive architecture:
- Task-level recursion: contrastive Propose–Probe–Compose steps reliably improve the harness for the current task without a verifier, scaling in parallel or serially
- Harnessing-level recursion: experience distilled from these modifications continuously updates an "improvement harness," letting the agent adapt better to future tasks
- Results: across 3 specialized benchmarks and 4 model-harness configurations, a single step already yields consistent gains for GLM-5 and Gemini 3.5 Flash—up to +11.5 on WorkBuddy-Bench, +4.5 on JobBench, +11.8 percentage points on LAB—with further gains as task-level recursion budget grows
More from coding & agent
- GraphicalUI impresses: AI design system tool auto-tidies your messy component soup — ow · 2026-10-14
- XERJ Corpus Hub: OSS reference memory that makes coding agents write Rust like a senior dev — HeyAmit_ · 2026-10-14
- One prompt turns Grok bots into a self-directing team that fixes your life admin — elonmusk · 2026-10-14
- The original 'chief of staff' prompt that puts all your bots to work — elonmusk · 2026-10-14
- Open Intelligence Summit: Nous Research on running serious agents without a single model provider — HeyAmit_ · 2026-10-14
- Multiplayer Grok bot over email: bot-to-bot debugging, support and deals — elonmusk · 2026-10-14