Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Google has released Harness², an open-source framework that improves an AI agent’s surrounding software—its prompts, tools, memory, skills and orchestration—while leaving the underlying model weights frozen. It recursively tests harness changes on tasks the system has not seen during harness design, then carries useful lessons into future tasks.
What Harness² is
The official Google Research repository defines a harness as the system around a model that enables it to act as an agent: prompts, tools, memory, skills, orchestration and related runtime code. Harness² applies recursive optimization to that layer in an open-world setting, where tasks are unseen when the initial harness is built.
This distinction matters. Harness² does not fine-tune, retrain or otherwise change the model’s neural-network weights. It changes the software context in which a frozen model operates. The same model can therefore behave differently after its system prompt, tool instructions, memory policy or orchestration logic has been refined.
Free tools Windows power users keep installed
One-click scans. No signup required.
The 2026 repository release and accompanying preprint describe two linked recursion levels. One improves execution of the task in front of the agent; the other learns how to improve future tasks.
#1 Best Overall
How the two recursion levels work
Task-level recursion refines the current task
For an unseen task, Harness² starts with a common execution harness and explores candidate edits. It runs those alternatives under matched conditions, compares what the agents actually did, and keeps changes supported by behavioral evidence. The process does not require a ground-truth answer or a task-specific label.
- Start with the current execution harness and an open-world task.
- Propose candidate edits to one or more harness components.
- Probe each candidate by executing it under comparable conditions.
- Compare the resulting behavior and compose the changes that show useful contrasts.
A candidate might produce better tool selection, clearer intermediate planning or fewer unnecessary actions. Harness² uses those observable differences—not a supplied answer key—to decide which edit is worth adopting.
Harnessing-level recursion learns for later tasks
After a task-level refinement, the system records the adopted harness and evidence-backed lessons in a persistent improvement library. Future proposals can draw on what previously worked, where it worked and why. This second loop updates the improvement mechanism itself rather than only the temporary runtime configuration for one task.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Parallel and sequential exploration
Harness² supports parallel recursion, which evaluates multiple candidates concurrently, and sequential recursion, which applies revisions one after another. In the tested settings, increasing the task-level budget improved results in both modes. Parallel execution can shorten elapsed time when sufficient concurrency is available, but it still performs more total work and therefore raises aggregate test-time computation.
Which parts of an agent can be changed?
The framework exposes eight editable runtime components. A deployment can refine one component or compose changes across several of them.
| Editable component | What it controls |
|---|---|
| System prompt | The agent’s role, priorities and general operating instructions. |
| Guardrails | Rules that constrain behavior, tool use and safety-sensitive actions. |
| Memory | What information is retained, retrieved and presented to the model. |
| Tool guidance | Instructions for selecting, calling and interpreting external tools. |
| Subagents | Delegation to specialized agents and the way their results are combined. |
| Skills | Reusable capabilities or procedures available to the agent. |
| Scripts | Executable helper logic used during task completion. |
| Middleware plugin | Runtime interception and coordination between the model, tools and other components. |
Why Propose–Probe–Compose is different
Conventional optimization often needs labeled examples, a verifier or a known correct answer. Harness² uses a contrastive Propose–Probe–Compose procedure instead. It proposes edits, probes their effects under matched execution conditions, and composes changes that are supported by contrasts in behavior.
The improvement library gives this method continuity across tasks. An adopted harness is not merely a one-off prompt variant; it becomes evidence that can influence later proposals. Removing that library in the paper’s ablation reduced WorkBuddy-Bench by 3.1 points and JobBench by 1.8 points. Removing the system’s scoping mechanism reduced those benchmarks by 2.9 and 3.7 points, respectively, while removing composition reduced them by 1.6 and 1.9 points.
Benchmark results reported for the 2026 release
The repository and paper evaluate four executor–harness configurations across WorkBuddy-Bench, JobBench and Harvey’s Legal Agent Benchmark (LAB). The figures below are maximum improvements after one Harness² step; they are benchmark measurements, not guarantees for every workload.
| Evaluation | Reported improvement | Qualification |
|---|---|---|
| WorkBuddy-Bench | Up to 11.5 points | Maximum gain after one Harness² step. |
| JobBench | Up to 4.5 points | Maximum gain after one Harness² step. |
| LAB partial pass rate | Up to 11.8 percentage points | Maximum gain after one Harness² step. |
| Parallel recursion on WorkBuddy-Bench | Up to 3.9 points | Additional gain reported in the paper’s parallel-recursion analysis. |
| Parallel recursion on JobBench | Up to 4.2 points | Additional gain reported in the paper’s parallel-recursion analysis. |
| Parallel recursion on LAB | Up to 6.7 percentage points | Additional partial-pass-rate gain reported in the paper’s analysis. |
With recursion budget k=5, Gemini 3.5 Flash running under OpenCode gained as much as 7.5 points on JobBench and 8.4 points on WorkBuddy-Bench relative to its base configuration.
Transfer is possible but not uniform. A harness refined for Gemini 3.5 Flash improved GLM-5 by 17.3 points on the Code evaluation. The paper reports that transfer gains for Gemini 3.7 Flash and Opus 4.8 varied by domain, so a harness that helps one model or task family should not be assumed to work equally well elsewhere.
Compute costs and operational trade-offs
Recursive improvement consumes additional test-time computation. Each candidate requires executions, and the editors also use tokens to propose and assess changes. A larger recursion budget can improve scores, but it increases execution and editor-token costs. Parallel mode may reduce wall-clock latency only when the environment can run candidates concurrently; it does not remove the underlying work.
Because the model weights remain frozen, Harness² is best understood as runtime optimization rather than model training. Teams can iterate on prompts, tools and orchestration without maintaining a new model checkpoint, but they must budget for repeated evaluations and monitor whether an edit helps the target domain without harming another.
Best Value
Installation and running the demos
Local demonstration
The repository’s local demo is designed to run without model credentials. Its stated baseline is Linux, Python 3.12, the uv environment manager and the harness2-demo entry point. The demo uses deterministic editors and real Python executions, and it exercises both parallel and sequential recursion modes.
- Use a Linux environment with Python 3.12.
- Install and configure
uvaccording to the project’s repository instructions. - Set up the Harness² checkout and its Python environment with the project’s supplied configuration.
- Run the provided
harness2-democommand to execute the credential-free demonstration.
Model-backed benchmark runs
Running the published benchmark configurations requires more infrastructure than the local demo:
- Node.js 20 or newer.
- Google Cloud CLI authentication and configured model IDs.
- Dependencies specific to each benchmark.
- Pandoc and Poppler for LAB.
- Docker for WorkBuddy-Bench.
Exact model identifiers, executor settings and benchmark setup should be taken from the version of the repository being used, because those details determine both the measured score and the compute bill.
Recommended Free Tools
Status and limitations
The numbers come from a 2026 preprint and repository release. They establish that recursive harness changes can improve the tested configurations; they do not establish a universal gain for every agent, model, domain or production workload. Transfer varies, and additional executions and editor tokens are part of the cost.
The repository states that Harness² is not an officially supported Google product and is not eligible for Google’s Open Source Software Vulnerability Rewards Program. Those status terms should be considered before using it in a production security process.
For teams evaluating agent quality without retraining a model, Harness² offers a concrete way to search the surrounding runtime: generate alternatives, test them against matched behavior, retain supported edits and build a library of lessons for the next task.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →
