Google 开源递归智能体框架 Harness²

Add as a preferredsource on Google

Google Cloud AI Research has open-sourced Harness². It is a recursive agent harnessing framework that improves the code and files around a frozen model. For developers, the pitch is simple. Your agent may get better without a better model.

TL;DR

Size: A framework, not a model. 0 parameters trained; 8 editable harness components. Version 0.1.0.

Runs on: Linux with Python 3.12 or 3.13 and uv. Model runs call models through Vertex AI. GPU requirement: Not disclosed.

Performance: 1 refinement step improved all 4 model and harness setups on all 3 benchmarks.

Best: GLM-5 under OpenCode rose from 64.2 to 75.7 on WorkBuddy-Bench (+11.5 points).

Worst: GLM-5 under Codex gained only 1.2 points on JobBench (23.6 to 24.8).

What is Harness²?

Harness² is a plug-and-play framework from Google Research that refines an agent’s harness during deployment. A harness is everything around the model: prompts, tools, memory, skills and orchestration. Harnessing is the process of optimizing that system. Model weights never change.

The target setting is the open world. Tasks arrive unseen, and no verifier says whether an answer is right. The research team argues that most self-improvement methods assume labels or test cases. Harness² is built to work without them.

What are the 8 harness components?

Harness² exposes 8 editable components, each mounted as a file:

How does recursive agent harnessing work?

The framework runs two linked levels of recursion:

Task-level recursion uses a contrastive Propose, Probe, Compose step. First, the agent runs the task once with the base harness. An editor then proposes 2 candidates. The invariant candidate targets general agent failures such as planning or tool use. The domain candidate targets knowledge and procedures for the current domain. Both candidates run on the same task. The composer compares their trajectories against the reference and keeps only edits backed by observed behavior.

Harnessing-level recursion happens after each task. A reflector stores the adopted harness plus an EXPERIENCE.md file of lessons. This library guides future proposals. Each new task still starts from the base harness.

The budget k scales the search. Parallel mode probes k invariant and k domain candidates, then makes k compositions. Sequential mode chains k refinement steps.

How does Harness² perform on benchmarks?

The research team tested 3 professional benchmarks: WorkBuddy-Bench, JobBench and Harvey’s LAB. Executors were Gemini 3.5 Flash and GLM-5 under OpenCode and Codex. The default editor was Claude Opus 4.8. Results are means over 3 runs.

Setup (single step, k=1)WorkBuddy AvgJobBenchLAB Partial

Gemini 3.5 Flash, OpenCode72.4 → 77.531.5 → 36.079.9 → 81.9

Gemini 3.5 Flash, Codex74.5 → 77.533.9 → 35.371.6 → 75.6

GLM-5, OpenCode64.2 → 75.720.6 → 24.259.6 → 71.4

GLM-5, Codex64.4 → 75.323.6 → 24.858.1 → 65.9

Source: Harness² paper, Tables 1 and 2.

Weaker executors gained more. After 1 step, GLM-5 beat base Gemini 3.5 Flash on WorkBuddy-Bench. Parallel recursion at k=5 lifted Gemini under OpenCode to 80.8 on WorkBuddy. On WorkBuddy-Bench, 3 of 4 setups also used fewer steps per task. Gemini under OpenCode fell from 25.4 to 21.9.

How does it compare with other harness methods?

FeatureHarness²Meta-HarnessReasoningBankHarness Scaling

DevelopersGoogle Cloud AI Research, NorthwesternStanford IRIS Lab and co-authorsGoogle Research and co-authorsWang et al.

What changesFull 8-component harness, per taskHarness code, searched offlineRetrieved memory; harness fixedHarness, revised per instance

Needs training splitNoYesNoNo

Learns across tasksYes (library + lessons)Deploys 1 best harnessYes (memory)No

Code licenseApache-2.0MITNot disclosedNot disclosed

WorkBuddy Avg, Gemini/OpenCode*77.5 (80.8 at k=5)76.473.475.3

JobBench, Gemini/OpenCode*36.035.830.832.9

LAB Partial, Gemini/Codex*75.676.574.867.9

*Scores as measured in the Harness² paper; the other methods did not report these benchmarks themselves.

How do you install and run Harness²?

The README targets Linux, Python 3.12 and uv. A local demo runs both recursion modes without model credentials: Copy CodeCopiedUse a different Browser

uv venv --python 3.12 .venv source .venv/bin/activate uv pip install -e '.[render]' harness2-demo --output runs/demo --k 2

Model-backed runs need Node.js 20+, the Google Cloud CLI and a Vertex AI project. LAB needs Pandoc and Poppler. WorkBuddy-Bench needs Docker. Executors are OpenCode by default, or Codex with --substrate cx.

Key Takeaways

Harness² rewrites an agent’s harness at test time; weights stay frozen.

1 step lifted GLM-5 by 11.5 points on WorkBuddy-Bench.

Propose, Probe, Compose replaces ground-truth labels with behavioral contrasts.

A lessons library carries improvements to future tasks.

Expect extra runs per task, and sandbox it yourself.

Asif RazzaqWebsite | + postsBio ⮌

Asif RazzaqMeet Voxel51 3.0: A Unified Multimodal Data Platform for Physical AI

Asif RazzaqMeet Kyutai REKI: The 2.4B VLA Model Bringing Efficient Memory to Robot Manipulation

Asif RazzaqDeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API

RELATED ARTICLESMORE FROM AUTHOR

Meet Voxel51 3.0: A Unified Multimodal Data Platform for Physical AI

Meet Kyutai REKI: The 2.4B VLA Model Bringing Efficient Memory to Robot Manipulation

DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API