Google Open-Sources Harness2, a Recursive Agent Harnessing Framework

Google Cloud AI Analysis has open-sourced Harness². It’s a recursive agent harnessing framework that improves the code and information round a frozen mannequin. For builders, the pitch is straightforward. Your agent might get higher with out a higher mannequin.

TL;DR

Dimension: A framework, not a mannequin. 0 parameters skilled; 8 editable harness elements. Model 0.1.0.
Runs on: Linux with Python 3.12 or 3.13 and uv. Mannequin runs name fashions by way of Vertex AI. GPU requirement: Not disclosed.
Efficiency: 1 refinement step improved all 4 mannequin and harness setups on all 3 benchmarks.
Greatest: GLM-5 beneath OpenCode rose from 64.2 to 75.7 on WorkBuddy-Bench (+11.5 factors).
Worst: GLM-5 beneath Codex gained just one.2 factors on JobBench (23.6 to 24.8).

What’s Harness²?

Harness² is a plug-and-play framework from Google Analysis that refines an agent’s harness throughout deployment. A harness is every part across the mannequin: prompts, instruments, reminiscence, expertise and orchestration. Harnessing is the method of optimizing that system. Mannequin weights by no means change.

The goal setting is the open world. Duties arrive unseen, and no verifier says whether or not a solution is true. The analysis staff argues that almost all self-improvement strategies assume labels or check circumstances. Harness² is constructed to work with out them.

What are the 8 harness elements?

Harness² exposes 8 editable elements, every mounted as a file:

System immediate
Guardrails
Reminiscence
Instrument steerage
Subagents
Abilities
Scripts
Plugins (middleware)

They cowl context development, the instrument interface, delegation and habits contained in the instrument loop. The format extends the 7-file break up from earlier Agentic Harness Engineering work by 1 element.

How does recursive agent harnessing work?

The framework runs two linked ranges of recursion:

Process-level recursion makes use of a contrastive Suggest, Probe, Compose step. First, the agent runs the duty as soon as with the bottom harness. An editor then proposes 2 candidates. The invariant candidate targets common agent failures resembling planning or instrument use. The area candidate targets information and procedures for the present area. Each candidates run on the identical process. The composer compares their trajectories towards the reference and retains solely edits backed by noticed habits.
Harnessing-level recursion occurs after every process. A reflector shops the adopted harness plus an EXPERIENCE.md file of classes. This library guides future proposals. Every new process nonetheless begins from the bottom harness.

The finances okay scales the search. Parallel mode probes okay invariant and okay area candidates, then makes okay compositions. Sequential mode chains okay refinement steps.

<div class="hd"><div class="dots"><i model="background:#4285F4"></i><i model="background:#EA4335"></i><i model="background:#FBBC04"></i><i model="background:#34A853"></i></div><h3>Harness<sup>2</sup> defined</h3></div>
<div class="sub">Google's recursive agent harnessing: the harness adjustments, the mannequin weights keep frozen.</div>

<div class="tabs">
<button class="tab on" data-t="p1">The way it works</button>
<button class="tab" data-t="p2">8 elements</button>
<button class="tab" data-t="p3">Benchmarks</button>
<button class="tab" data-t="p4">vs baselines</button>
</div>

<div class="pn on" id="h2-p1">
<div class="lvl">01 · Process-level recursion (inside 1 process)</div>
<div class="movement">
<div class="nd" model="–c:#9aa0a6"><span class="tg" model="background:#9aa0a6">REFERENCE</span><b>Run base harness</b><small>H0 executes the duty as soon as</small></div>
<div class="nd" model="–c:#4285F4"><span class="tg" model="background:#4285F4">PROPOSE</span><b>2 candidates</b><small>Invariant + domain-specific edits</small></div>
<div class="nd" model="–c:#EA4335"><span class="tg" model="background:#EA4335">PROBE</span><b>Run each</b><small>Identical process, identical mannequin, evaluate habits</small></div>
<div class="nd" model="–c:#FBBC04"><span class="tg" model="background:#FBBC04">COMPOSE</span><b>Merge what labored</b><small>Preserve edits backed by proof</small></div>
</div>
<div class="lvl">02 · Harnessing-level recursion (throughout duties)</div>
<div class="movement">
<div class="nd" model="–c:#4285F4"><span class="tg" model="background:#4285F4">EXECUTE</span><b>Last harness</b><small>Deliverable returned</small></div>
<div class="nd" model="–c:#34A853"><span class="tg" model="background:#34A853">REFLECT</span><b>Evaluate the document</b><small>Which edits helped, and why</small></div>
<div class="nd" model="–c:#34A853"><span class="tg" model="background:#34A853">UPDATE</span><b>Enchancment library</b><small>Harness + EXPERIENCE.md</small></div>
<div class="nd" model="–c:#9aa0a6"><span class="tg" model="background:#9aa0a6">NEXT TASK</span><b>Restart at H0</b><small>Library guides new proposals</small></div>
</div>
<div class="be aware" id="h2-note">Press <b>Play</b> or <b>Step</b> to stroll by way of 1 Harness<sup>2</sup> enchancment step.</div>
<div class="ctl"><button class="btn pri" id="h2-play"> Play</button><button class="btn" id="h2-step">Step →</button><button class="btn" id="h2-reset">↺ Reset</button><div class="prog"><i id="h2-prog"></i></div></div>
<div class="fn">At okay=1: 1 reference + 2 probes + 1 remaining run per process. Parallel mode probes okay invariant and okay area candidates; sequential mode chains okay steps.</div>
</div>

<div class="pn" id="h2-p2">
<div class="cg" id="h2-cg"></div>
<div class="be aware" id="h2-cnote" model="border-left-color:#4285F4">Click on a element to see its runtime position.</div>
<div class="lg"><span model="–c:#4285F4">Context development</span><span model="–c:#EA4335">Instrument interface</span><span model="–c:#FBBC04">Orchestration</span><span model="–c:#34A853">Middleware</span></div>
</div>

<div class="pn" id="h2-p3">
<div class="sel" id="h2-sel"></div>
<div class="bm" id="h2-bm"></div>
<div class="fn">Single enchancment step (okay=1). Means over 3 runs. LAB = Partial cross price (%). Supply: Harness<sup>2</sup> paper, Tables 1 and a pair of.</div>
</div>

<div class="pn" id="h2-p4">
<div class="sel" id="h2-sel2"></div>
<div class="bx rk" id="h2-rk"></div>
<div class="fn">Gemini 3.5 Flash executor. Bar axis begins at 50. Meta-Harness is optimized offline on separate coaching duties. Scores as measured within the Harness<sup>2</sup> paper.</div>
</div>

<div class="ft"><span>Sources: arXiv 2610.14224 · github.com/google-research/harness2</span><a href="https://www.marktechpost.com" goal="_blank" rel="noopener">© Marktechpost</a></div>

(perform(){var f=doc.getElementById(‘mtp-h2x-frame’);window.addEventListener(‘message’,perform(e){if(e.supply===f.contentWindow&&e.knowledge&&e.knowledge.mtpH2xHeight){f.model.peak=e.knowledge.mtpH2xHeight+’px’;}});})();

How does Harness² carry out on benchmarks?

The analysis staff examined 3 skilled benchmarks: WorkBuddy-Bench, JobBench and Harvey’s LAB. Executors had been Gemini 3.5 Flash and GLM-5 beneath OpenCode and Codex. The default editor was Claude Opus 4.8. Outcomes are means over 3 runs.

Setup (single step, okay=1)	WorkBuddy Avg	JobBench	LAB Partial
Gemini 3.5 Flash, OpenCode	72.4 → 77.5	31.5 → 36.0	79.9 → 81.9
Gemini 3.5 Flash, Codex	74.5 → 77.5	33.9 → 35.3	71.6 → 75.6
GLM-5, OpenCode	64.2 → 75.7	20.6 → 24.2	59.6 → 71.4
GLM-5, Codex	64.4 → 75.3	23.6 → 24.8	58.1 → 65.9

Supply: Harness² paper, Tables 1 and a pair of.

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

How does it evaluate with different harness strategies?

Characteristic	Harness²	Meta-Harness	ReasoningBank	Harness Scaling
Builders	Google Cloud AI Analysis, Northwestern	Stanford IRIS Lab and co-authors	Google Analysis and co-authors	Wang et al.
What adjustments	Full 8-component harness, per process	Harness code, searched offline	Retrieved reminiscence; harness fastened	Harness, revised per occasion
Wants coaching break up	No	Sure	No	No
Learns throughout duties	Sure (library + classes)	Deploys 1 greatest harness	Sure (reminiscence)	No
Code license	Apache-2.0	MIT	Not disclosed	Not disclosed
WorkBuddy Avg, Gemini/OpenCode*	77.5 (80.8 at okay=5)	76.4	73.4	75.3
JobBench, Gemini/OpenCode*	36.0	35.8	30.8	32.9
LAB Partial, Gemini/Codex*	75.6	76.5	74.8	67.9

*Scores as measured within the Harness² paper; the opposite strategies didn’t report these benchmarks themselves.

How do you put in and run Harness²?

The README targets Linux, Python 3.12 and uv. An area demo runs each recursion modes with out mannequin credentials:

Copy CodeCopiedUse a distinct Browser

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

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

Key Takeaways

Harness² rewrites an agent’s harness at check time; weights keep frozen.
1 step lifted GLM-5 by 11.5 factors on WorkBuddy-Bench.
Suggest, Probe, Compose replaces ground-truth labels with behavioral contrasts.
A classes library carries enhancements to future duties.
Count on additional runs per process, and sandbox it your self.

Take a look at the Paper, GitHub Repo, Paper and Mission Web page. All credit score goes to the researcher of this mission. Additionally, be happy to observe us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Publication. Wait! are you on telegram? now you’ll be able to be part of us on telegram as nicely.