Google Cloud AI Research open-sources Harness²

Google Cloud AI Research open-sourced Harness², a framework for recursively refining agent harnesses while keeping model weights fixed.[1][2] Its target is work that was not seen when the harness was designed and may not have ground-truth verifiers; the framework’s edits do not require ground-truth feedback.[3][2] The repository lists Apache-2.0 as the project’s license.[2] MarkTechPost’s report shows the date October 18, 2026.[1]

Harness² connects two levels of recursion.[2] At the task level, it uses Propose–Probe–Compose: propose harness edits, test them, and compose a revision using observed behavior.[2] After a task, it updates a persistent improvement harness with adopted harnesses and evidence-supported lessons for future refinement.[2] The editable components include the system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins.[2] Together, these components cover context construction, the tool interface, delegation, and behavior inside the tool loop.[2]

The paper reports results on WorkBuddy-Bench, JobBench, and Harvey’s Legal Agent Benchmark.[3] In the reported experiments, Gemini 3.5 Flash and GLM-5 served as executor models under OpenCode and Codex configurations, while Claude Opus 4.8 was the default editor model.[1] A single step improved both executor models over their base harnesses.[2] The largest reported single-step WorkBuddy-Bench gain was 11.5 points: GLM-5 under OpenCode rose from 64.2 to 75.7.[1] The paper also reports maximum single-step gains of 4.5 points on JobBench and 11.8 percentage points on LAB.[3] Results are means over three runs.[1]

The release supports sequential and parallel recursion, OpenCode and Codex executors, and all three named benchmarks.[2] With parallel recursion at k=5, Gemini under OpenCode scored 80.8 on WorkBuddy-Bench.[1] MarkTechPost reports that weaker executor models gained more from a single refinement step.[1] The local demo can run both recursion modes without model credentials.[2] Model-backed benchmark runs require Node.js 20+, the Google Cloud CLI, and a Vertex AI project, and the README warns that these runs incur API costs.[1][2]