Google's Harness² and the Free Harness Engineering Course, Checked
Google Research open-sourced Harness², which reworks an agent's setup for each task, and a free course teaches you to build that setup by hand. We checked the repo, the paper and the course: what is real, what the post leaves out, where kits fit.
On October 18, 2026 Akshay (@akshay_pachaar) posted that "Google just open-sourced Harness²", a framework that lets an agent rework its own setup for every new task while the model stays the same. Two days earlier Tech with Mak (@techNmak) shared Walking Labs' free course Learn Harness Engineering, which teaches you to build that setup by hand.
Both posts are about the same thing: the harness, everything around the model that turns it into a working agent. We checked both against their repositories, license files and the Harness² paper. The repository is real Google Research code under Apache-2.0, but it is research software, and the post leaves out who edits the harness, who grades the results and what the extra runs cost. The course is what the post says it is.
The original post by @akshay_pachaar on X (October 18, 2026)
The post about the course by @techNmak on X (October 16, 2026)
What a harness is
The model writes the answer; the harness decides what the model sees, which tools it may call, what it remembers and when the work counts as done. If you use Claude Code, you already edit one: CLAUDE.md, skills, subagents, hooks and MCP servers are all harness. The two projects split it up a little differently:
Is Harness² really from Google?
Yes, with a qualifier. The repository sits in the google-research organization on GitHub, the one named Google Research that links to research.google. Its LICENSE file is the Apache License 2.0, so you can use, change and ship it commercially. The paper, arXiv 2610.14224, was submitted on October 9, 2026, and its author list names Google Cloud AI Research and Northwestern University as affiliations.
The qualifier is the README's last line: the project calls itself not an officially supported Google product. It is research code from Google researchers, not a Google Cloud service or a Gemini feature. The repository was created on October 6, 2026 and had 4,961 stars when we read it; 2,925 of them arrived on October 18, the day of the post.
How Harness² works
For each task, the agent first runs with its current harness. Then an editor proposes changes, the changed harnesses are tried on the same task, and a composer builds a new harness from what went differently between the runs. No answer key is needed for that step. After the task, the adopted harness and the lessons that held up go into a library that the next task starts from.
The --k option sets the budget. In parallel mode it tries several candidates at once (wider); in sequential mode it refines one harness step by step (deeper). The agent doing the work runs in OpenCode by default, or in Codex. Setup asks for Linux, Python 3.12 and, for the model-backed benchmarks, a Vertex AI project; a demo runs without any model keys.
What the paper reports
All results below are as reported by the authors; nobody outside has repeated them yet. They tested two models, Gemini 3.5 Flash and GLM-5, each inside OpenCode and Codex, on three benchmarks: WorkBuddy-Bench (code, web and office tasks), JobBench (professional deliverables) and LAB, Harvey's legal benchmark.
| Benchmark, OpenCode | Model | Base | One step |
|---|---|---|---|
| JobBench score | Gemini 3.5 Flash | 31.5 | 36.0 |
| JobBench score | GLM-5 | 20.6 | 24.2 |
| LAB criterion pass rate | Gemini 3.5 Flash | 79.9 | 81.9 |
| LAB criterion pass rate | GLM-5 | 59.6 | 71.4 |
The weaker model gains most: by our calculation one step adds 11.5 points to GLM-5 in OpenCode on WorkBuddy-Bench and 5.1 points to Gemini 3.5 Flash. More budget helps further: with parallel recursion at k=5 the paper reports up to 3.9 more points on WorkBuddy-Bench. In an appendix the authors also ran Gemini 3.5 Flash in four harnesses left as they ship, including Claude Code, which was the best of them on code and office tasks; Harness², starting from plain OpenCode, beat the best fixed one by 4.6, 5.8 and 4.9 points on code, web and office.
What the post gets right
- The models, the three kinds of work (legal, office, professional), the single step that beats the original harness and the gains from more rounds all match the paper.
- Parallel and sequential recursion, OpenCode and Codex as executors, and the library that carries lessons to the next task are all in the README.
- Refining a harness does not need a correct answer: the editor compares runs of the same task.
What the post leaves out
- A different model does the editing. "The agent proposes edits to its harness" is loose. In the main results the harness editor is Claude Opus 4.8 running in OpenCode, a stronger model than the two being improved. The paper does test the same model in both roles, and reports that all five models it tried improved, so a stronger editor is not required; but that is not the setup behind the headline numbers.
- The scores come from AI judges. No grader is needed to refine the harness, but the reported results are graded: JobBench deliverables by Claude Opus 4.8 against a rubric, LAB answers by Claude Sonnet 4.6.
- It costs more runs. Every step runs the task several times plus editor sessions, and the authors list the added inference cost as a limitation. The README warns that model-backed runs cost money and suggests starting small.
- It can pick the worse version. The paper's failure analysis shows cases where selection preferred a weaker candidate and where a wrong assumption spread across all candidates.
- Run it in a sandbox. The README says the executor has unrestricted network access and Codex runs with its sandbox disabled.
The course: Learn Harness Engineering
The second post checks out. Walking Labs' course has 14 lectures and 8 hands-on projects, and all projects build on the same Electron app, each one's solution becoming the next one's starter. It is MIT-licensed, free, and offered in 15 languages. You need Claude Code, Codex or a similar agent that can edit files and run commands.
| Phase | Lectures and projects |
|---|---|
| See the problem | Why strong models still fail; what a harness is; prompt-only vs. rules-first (P01) |
| Structure the repo | The repo as the single source of truth; split instructions across files |
| Connect sessions | Progress logs and hand-offs; initialize before each session |
| Feedback and scope | One feature at a time; feature lists; runtime feedback |
| Verification | Stop agents from declaring victory early; end-to-end runs |
| Put it together | Observability, clean hand-off, the capstone harness (P06) |
| Loops and graphs | Automated loops with a maker and a checker; workflows drawn as graphs |
Our summary of the course's learning path, from its README.
It also ships copy-ready templates (AGENTS.md, feature_list.json, init.sh), a harness-creator skill that scaffolds a harness for your own project, and audit-harness.sh, a shell script that checks a repo against the five subsystems. A newer section takes apart the harnesses of Claude Code, Codex, Pi and DeepSeek. Two small things: one section of the README still says 13 lectures and 7 projects, from before the last two were added, and the audit script is shown piped from the web straight into bash, so read it before you run it.
The repository was created on March 29, 2026 and had 20,723 stars when we read it; its busiest day was October 1, 2026, with 794 new stars.
Where AgentGid's starter kits fit
There are three ways to get a good harness: build it yourself (the course), let a model refine it per task (Harness²), or start from one that has already been tested on your kind of project. The third is what an AgentGid starter kit is. Each kit writes the harness files your agent builds from: an AGENTS.md with the plan a step at a time and a check after each step, the known traps that earlier test builds ran into, a definition of done, and the MCP setup for Claude Code, Cursor, Codex or Copilot. The MCP server kit and the SaaS starter kit are good examples.
In the course's terms a kit covers instructions, scope and verification for one project type. In Harness² terms it is the kind of adopted harness and lessons its library builds up, written down and checked in advance.
Harness² also echoes the Claude Code advisor: a stronger model is called in to improve how a cheaper one works. Our verdict there applies here too. Much of what the stronger model is paid to find (a plan that needs checking, a loop of repeated errors, a task called done too early) can be written into the harness before the first prompt.
The post ends by pointing to the author's earlier piece on Comet's Opik, which fixes harnesses from failing traces in production; that is a separate project.
The verdict
- Harness² is genuine Google Research code under Apache-2.0, not a Google product. Read it as a paper with working code: the gains are as reported on three benchmarks, with a Claude model editing and Claude models judging.
- Not for your daily coding agent yet. It runs benchmark tasks with extra runs per task; the useful takeaway now is the idea of comparing runs of the same task before changing your setup.
- The course is the best free way to learn to do it by hand, with templates you can copy into a repo today.
- For a new project, start from a kit and change the harness from there.
Facts about Harness² come from its repository, LICENSE and the arXiv paper; facts about the course from its README and LICENSE; star counts from GitHub's API on October 18, 2026. Score gains are by our calculation from the paper's tables. We did not run Harness² or take the course.