Harness2 by Google Research – Recursive Agent Harnessing

Harness² is a plug-and-play framework for recursive agent harnessing that refines how language models act as agents without modifying model weights. It implements two nested recursions: a task-level PROPOSE-PROBE-COMPOSE loop that proposes harness edits, probes their behavior on the same task, and composes a refined harness from behavioral contrasts; and a harnessing-level recursion that consolidates adopted harnesses and evidence into a persistent improvement harness to guide future tasks. Harness components are explicitly editable - system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins - so context construction, tool interfaces, delegation, and runtime behavior can be changed and re-evaluated. The pipeline supports parallel and sequential task-level recursion, OpenCode and Codex executors, and both sandboxed demos and model-backed runs.

Benchmarks and experiments show concrete gains: a single Harness² step improved GLM-5 and Gemini 3.5 Flash over base harnesses with up to +11.5 points on WorkBuddy-Bench, +4.5 on JobBench, and +11.8 percentage points on LAB’s partial pass rate, with larger gains when increasing the recursion budget. Installation targets Linux and Python 3.12 (local demo requires no model credentials), while model-backed runs need Node.js, Google Cloud CLI, and benchmark-specific deps. Results, adopted harnesses, reflections, and evaluations persist under runs/ for replay and resumption. Harness² is open-source under Apache 2.0 and includes guides, benchmarks (JobBench, LAB, WorkBuddy), and reproducible experiment recipes.