Harness2
An agent framework open-sourced by a Google research team that recursively rewrites the agent harness structure during the testing phase, improving model performance without updating model weights. It is released under the Apache-2.0 license.
Core Facts
Timeline (last 90 days)
From a single link on Hacker News alone, it cannot be confirmed whether Harness2 is a paper, an open-source code repository, or a technical blog.
Judging from Google Research's release channels, Harness2 is more likely a research-oriented prototype or conceptual framework rather than a ready-to-use product.
Harness2's working loop is: the Agent proposes modifications to the existing Harness for a new task, runs several candidate versions in parallel, and selects the better one by comparing actual performance.
Google Research has open-sourced Harness2, a framework that gives AI Agents' "scaffolding" the ability to self-improve.
Throughout Harness2's improvement process the model weights remain unchanged; the improvement happens entirely at the configuration layer.
After a Harness2 task is completed, the improved Harness and the related experience are saved, so that subsequent edits start from a higher baseline.
Harness2 uses grader-free relative evaluation, selecting the best through lateral comparison among candidate versions, without needing a ground-truth answer or an external grader.
Harness2 and model fine-tuning are complementary: fine-tuning changes the model itself, while Harness2 changes how the model is used, and the latter is lighter and easier to transfer across models.
Google validated Harness2 on two models, Gemini 3.5 Flash and GLM-5, covering three types of benchmarks: legal, office, and professional work.
In the benchmark tests, even a single round of Harness editing already outperformed the original configuration; as the number of editing rounds increased, scores continued to rise.
12 more timeline events
All Facts (20)
The core idea of Harness2 is to recursively rewrite the agent harness at test time without updating the model weights.
65%UnverifiedFrom a single link on Hacker News alone, it cannot be confirmed whether Harness2 is a paper, an open-source code repository, or a technical blog.
50%UnverifiedIn the benchmark tests, even a single round of Harness editing already outperformed the original configuration; as the number of editing rounds increased, scores continued to rise.
50%UnverifiedHarness2's working loop is: the Agent proposes modifications to the existing Harness for a new task, runs several candidate versions in parallel, and selects the better one by comparing actual performance.
50%UnverifiedThroughout Harness2's improvement process the model weights remain unchanged; the improvement happens entirely at the configuration layer.
50%UnverifiedAfter a Harness2 task is completed, the improved Harness and the related experience are saved, so that subsequent edits start from a higher baseline.
50%UnverifiedHarness2 uses grader-free relative evaluation, selecting the best through lateral comparison among candidate versions, without needing a ground-truth answer or an external grader.
50%UnverifiedHarness2 and model fine-tuning are complementary: fine-tuning changes the model itself, while Harness2 changes how the model is used, and the latter is lighter and easier to transfer across models.
50%UnverifiedGoogle validated Harness2 on two models, Gemini 3.5 Flash and GLM-5, covering three types of benchmarks: legal, office, and professional work.
50%UnverifiedCurrently available public information only mentions "higher scores" and does not give the specific magnitude of improvement or benchmark details.
50%UnverifiedTesting with two models from different families, Gemini 3.5 Flash and GLM-5, shows that the Harness2 method itself does not depend on specific model characteristics.
50%UnverifiedAccording to official disclosures, a single refinement step alone improved GLM-5 by 11.5 points on WorkBuddy-Bench.
50%UnverifiedThe recursive rewriting of Harness2 means feeding the harness rewrite result back as the input for the next round of rewriting, forming an iteration.
50%UnverifiedRecursive rewriting requires the framework to have the ability to evaluate its own behavior; otherwise it will diverge or fall into a local optimum.
50%UnverifiedHarness2 extends test-time compute from searching over answers to searching over harness structures.
50%UnverifiedWhether the recursively rewritten harness produces behavioral drift across tasks requires longer-term evaluation data.
50%UnverifiedThe evaluation compute cost of each recursion in Harness2 was not quantified in the official disclosure.
50%UnverifiedThe discovery of this vulnerability first appeared on Hacker News, with a discussion heat of 5 points and 0 comments.
50%UnverifiedJudging from Google Research's release channels, Harness2 is more likely a research-oriented prototype or conceptual framework rather than a ready-to-use product.
50%UnverifiedGoogle Research has open-sourced Harness2, a framework that gives AI Agents' "scaffolding" the ability to self-improve.
50%