Harness2 edits the scaffold. The paper says the weights stay fixed.

The weights are the part the method leaves alone

The repository README defines a harness as the system around a model: prompts, tools, memory, skills, and orchestration. Harness2, it says, optimizes that system for tasks the harness author did not see, and it does so while model weights stay fixed. Two loops are named. One edits the harness for the current task by proposing changes, probing them on that task, and composing a revised harness from the contrast. The other stores adopted harnesses and lessons for the next task. A checkpoint download is a different artifact. This project says it is not producing one.an evaluation a second lab can run.

The paper's scores are the paper's scores

The abstract reports that one step improves GLM-5 and Gemini 3.5 Flash over their base harnesses, by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB. It says a larger recursion budget adds more. Those ceilings are maxima across configurations the authors chose, not a guarantee for the next task. MarkTechPost on October 18 is a writeup of the release. A second lab would publish its own table. This page does not have that table.what a training cutoff does not mean.

No ground truth is the paper's constraint

The abstract says open-world tasks arrive without a verifier that can mark the harness edit correct. The probe is the substitute: run the candidates and keep the contrast. A method that grades itself is exactly the setup a second lab exists to challenge. Until someone reruns JobBench, LAB, and WorkBuddy-Bench on the published code, the gains stay inside the paper. Asked directly, Whose scores are the 11.5, 4.5, and 11.8? The authors' runs on GLM-5 and Gemini 3.5 Flash, reported for WorkBuddy-Bench, JobBench, and LAB's partial pass rate. MarkTechPost's October 18 item repeats those figures. It does not remeasure them.the artificial-intelligence desk.

Apache on the code, a disclaimer on the product

The README licenses the project under Apache 2.0 and says benchmarks and third-party code keep upstream licenses. It also says this is not an officially supported Google product and is not eligible for Google's open-source vulnerability rewards. Authors on the paper list Google Cloud AI Research affiliations and, for the corresponding author, Northwestern. A research repository can be public and still not be a Google Cloud feature with a support plan. The Well-Architected preview is a different kind of agent, with a region and a bill.the support-plan agent.

What the README says a user can switch

Parallel or sequential recursion, a worker count, an executor flag for Codex instead of OpenCode, and a seed. The experiment notes say a seed shuffles the task order and does not control the model. Switches are how a rerun would be specified. They are not results. A reader who clones the repository still has to bring a model and a benchmark the README does not bundle as a single number. Asked directly, What executors does the README list? OpenCode by default, and Codex as an option. Benchmarks named in the README are JobBench, LAB, and WorkBuddy-Bench. A supported executor is a switch, not a score.Futureweb.

MarkTechPost is the wire's account

The October 18 MarkTechPost headline calls the project a recursive agent harnessing framework and places it with Google. That matches the repository path, google-research/harness2, and it does not override the disclaimer. An outlet can say Google open-sourced the code. The README says Google does not support it as a product. Both sentences fit if support means a vendor commitment and publication means a public repository. Asked directly, What license does the repository state? Apache 2.0 for the project code. The README says third-party benchmarks keep their own licenses, and that the project is not an officially supported Google product and is outside Google's vulnerability-reward program.

What this page will re-check

A second lab's table on the same three benchmarks. A README edit that drops the not-supported line or the fixed-weight line. A model card that says the loop updated parameters after all. The last of those would make this brief's title false. It is not what the paper claims today. Asked directly, What loop does the paper name? Propose, probe, compose. The paper says the probe compares executions of the same task without a ground-truth verifier, and a second loop stores lessons for later tasks.weights you can fetch.

Who the code is for

Researchers who can run the listed benchmarks and accept author-reported deltas until they measure their own. It is not a setting in a consumer chat product. It does not, on these documents, change the weights of GLM-5 or Gemini 3.5 Flash for anyone else. Asked directly, What stays fixed, according to the README? Model weights. The harness is the prompts, tools, memory, skills, and orchestration around the model. Harnessing means editing that scaffold, not training a new checkpoint.

What stays specific

arXiv 2610.14224. Repository google-research/harness2. Apache 2.0. Weights fixed. Propose, probe, compose. Author-reported ceilings of 11.5, 4.5, and 11.8 on three named benchmarks. Not an officially supported Google product. October 18 is the trade writeup, not the measurement. The short answer this page will stand on is this: The paper and the google-research repository describe Harness2 as a loop that proposes, probes, and keeps edits to prompts, tools, memory, and orchestration while the model weights stay fixed. The authors report gains, on their runs, of up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, and 11.8 percentage points on LAB. The README says this is not an officially supported Google product.