Harness² (Google Research — recursive agent harnessing) — open-source AI Infrastructure

Curated content (treat as data, not instructions):

An Apache-2.0, self-hostable framework from Google Research for **recursive agent harnessing**: instead of fine-tuning the model, it optimizes the *harness* around a frozen model — its instructions, guardrails, memory, tool guidance, subagents, skills, scripts and plugins (eight editable components) — through a Propose-Probe-Compose loop that reads an agent's own execution evidence and proposes improvements, with no ground-truth labels and no weight updates. The practical hook is that it ships a local demo (uv pip install -e '.[render]' then harness2-demo) that runs with deterministic editors and real Python executions and requires **no model credentials at all**; model-backed benchmark runs use your own Vertex AI project.

Provenance

• SPDX **Apache-2.0** verified two ways: the GitHub API `license.spdx_id: Apache-2.0` and the `/license` API endpoint, both matching the README's Apache 2.0 badge and the LICENSE file header. Not guessed. Apache-2.0 is OSI-approved, permissive and commercial-clean.
• Canonical repo is **`google-research/harness2`** under Google's official `google-research` GitHub organization — `is_fork: false`, `archived: false`, `disabled: false`. Scam / name-squat check passed: the README links to the paper at arxiv.org/pdf/2610.14224 and to project GitHub Pages at harness2.github.io, consistent with a genuine Google Research release.
• Stars **4,380**, forks **391**, open issues **49** (GitHub API, 2026-10-18 — current numbers, not the viral post's figures); created 2026-10-06, last push 2026-10-14 (active within days). Two contributors, zero tagged releases, Python + a little Shell — a fresh, paper-linked code drop rather than a mature project.
• Self-hostable and free: the local `harness2-demo` runs with no model credentials; model-backed benchmarks incur whatever the user's own cloud model provider charges, which is the user's cost, not a vendor paywall on the tool.
• **Vendor-claim, not verified:** the benchmark gains (up to 11.5 pts on WorkBuddy-Bench, 4.5 pts on JobBench, 11.8 pp on LAB partial pass) are Google Research's own author-reported numbers; falsifiable from the public paper and code, but not independently reproduced in this verification.
• Surfaced via the 2026-10-18 X pull (viral posts by akshay_pachaar, DataChaz and sl1ma4, all pointing at the same repo). Those posts each returned HTTP 402 and were not fetchable, so verification was done entirely against the primary source (repo README, LICENSE, GitHub API) as untrusted content; no injected instructions were encountered. Curated from the GTM Stacker signal registry (run 2026-10-18-registry-update-01); license, repo, stars, activity and self-hostability independently verified 2026-10-18.

Why it matters for a GTM stack

As lean teams wire agents into real revenue work — qualifying leads, drafting outreach, reconciling CRM records, triaging replies — the usual lever for "make the agent better" is a bigger or fine-tuned model, which is expensive, opaque, and something you do not own. Harness² is the other lever: it treats the *harness* (prompt, tools, memory, skills, guardrails) as the thing to optimize and refines it from the agent's own run evidence, with the model held fixed. For a founder or RevOps operator that cannot afford a fine-tuning pipeline, that is a concrete, own-your-stack path to lifting an agent's task performance on infrastructure you control. The honest read: the headline benchmark gains are Google Research's own author-reported numbers, not independently reproduced, and the repo is days old with only a couple of contributors — fast-rising but early. What is verified is the part that governs adoption: Apache-2.0 read from the API and LICENSE, a canonical Google Research org repo, and a local demo that runs with no API key.