Harness²: Google Research's framework that lets an agent rewrite its own harness
Published October 6, Apache-2.0: Harness² refines the system around a frozen model - prompts, tools, memory, skills - through a contrastive Propose-Probe-Compose loop and a persistent improvement harness. 9,061 stars and 796 forks observed October 19. The demo needs no model credentials; real benchmarks want Vertex AI and cost API money.
Facts checked October 19, 2026
Google Research published Harness² on GitHub on October 6: a plug-and-play framework for "recursive agent harnessing", where an agent rewrites the system around a frozen model - prompts, tools, memory, skills - for each new task. The repo is Apache-2.0, which I read in its LICENSE file. When I checked on October 19 it showed 9,061 stars and 796 forks with 72 open issues, thirteen days after creation. The local demo needs no model credentials; real benchmarks require a Google Cloud project and incur API costs. I did not install or run it. Repository
What Harness² is
A harness is everything around the model that makes it act as an agent: prompts, tools, memory, skills, orchestration. Harness² treats that surrounding system as the thing to optimize while the model weights stay fixed, and it connects two levels of recursion. Task-level recursion refines the execution harness for the current task through a contrastive Propose-Probe-Compose step: propose edits, probe their effects on the same task, and compose a refined harness from behavioral contrasts between executions, without ground-truth feedback. Harnessing-level recursion updates a persistent improvement harness with adopted harnesses and evidence-supported lessons from each task, so refinements accumulate across tasks the system has never seen. The framework exposes eight editable components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins.
License, popularity, maintenance
The repository carries the Apache License 2.0, verified by reading the LICENSE file directly. The popularity signal is strong but very young: 9,061 stars and 796 forks observed on October 19, against a creation date of October 6, so treat the count as launch attention rather than settled adoption. The repo shows two named contributors so far, with top commits from rux001, and 72 open issues. It is a google-research repository, which usually means a research artifact with no support commitment rather than a maintained product; there is also an accompanying paper, arXiv 2610.14224.
Setup reality
The documented path is Linux, Python 3.12 and uv: create the venv, install with the render extra, and run harness2-demo, which exercises both recursion modes with deterministic editors and real Python executions and requires no model credentials. That is the honest way to look at the machinery. Model-backed benchmarks are a heavier lift: Node.js 20+, the Google Cloud CLI, a Vertex AI project and model IDs in a .env file, and gcloud application-default credentials. The LAB legal benchmark additionally needs Pandoc and Poppler; WorkBuddy-Bench needs Docker. The README warns that model-backed runs incur API costs and suggests starting with a small task selection; a --dry-run flag previews task selection without model calls.
Privacy and security notes
The demo runs locally with no credentials, so a first look sends nothing anywhere. Model-backed runs are different: tasks, harness edits and executions flow through your Vertex AI project, so whatever data a benchmark task contains goes to Google Cloud under your own account and bill. WorkBuddy-Bench runs in Docker, which contains execution but also means giving the framework a container runtime on your machine. A framework whose whole point is rewriting its own guardrails and subagents deserves extra care: read what a proposed harness edit does before letting it run against real credentials, and keep the improvement harness's accumulated "lessons" away from sensitive projects until you trust what it stores.
Why it made the digest
Most agent frameworks compete on what the model can call. Harness² competes on what the system can rewrite about itself, and it ships the unglamorous parts that make that claim checkable: a no-credentials demo, named benchmarks with a judge and report step, and an explicit cost warning. The open question I would watch is the contrastive step: composing edits from behavioral contrasts without ground truth is exactly where a framework can quietly learn to reward verbose or lucky runs. The repo gives you the tools to see that for yourself, which is more than most launches this month can say.