What is the difference between parallel and sequential recursion?

Parallel recursion explores k independent candidates per scope and composes them, with each composer reading all candidates and execution evidence. Sequential recursion carries the composed harness through k successive improvement steps. Scaling parallel recursion from k=1 to k=5 added +4.3 points on WB-Code.

Is Harness² open source?

The project page links to a code repository and a paper (arXiv:2610.14224) with a BibTeX entry, indicating the code is publicly available for research use.

How much does Harness² cost?

The page does not list any pricing — it is a research project with code and a paper rather than a commercial product. Usage costs would come from the inference runs needed for reference execution, candidate probes, and final harness execution.

Harness² — Harness the agent. Then harness its self-improvement

Harness² is a research project from Google Cloud AI Research and Northwestern University that introduces recursive agent harnessing: a method that improves an AI agent's performance by editing the harness around a frozen model rather than the model itself. It works at two levels — refining the execution harness within a task, and updating an improvement library that guides how future harnesses are refined. The method is documented in a paper (arXiv:2610.14224) with accompanying code and benchmarks showing score improvements on WorkBuddy and JobBench.

Key Features

Two-level recursion

Task-level recursion refines the execution harness for the current task (Propose → Probe → Compose), while harnessing-level recursion updates an improvement library with reusable harnesses and experience (including EXPERIENCE.md files) to guide future tasks.

Frozen model, editable harness

The agent is defined as Model + Harness. Model weights stay fixed; the harness — made of eight editable components (system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins) — is what gets improved.

Parallel and sequential recursion modes

Parallel recursion explores k independent candidates per scope and composes them; sequential recursion carries the composed harness through k successive improvement steps. Scaling k from 1 to 5 adds +4.3 points on WB-Code.

Component-level edits beyond prompts

Improvements can touch what enters the context, tool usage, orchestration, and middleware — e.g., a JobBench case added verification guardrails, a reconciliation skill, and precise tool guidance rather than just prompt tweaks.

Per-task reset to base harness

Each new task starts the executor at the base harness H₀; reusable knowledge lives in the improver's library rather than in the executor's harness, avoiding unbounded harness drift.

Benchmarked results

On WorkBuddy (Code, Web, Office domains), a single improvement step lifts Gemini 3.5 Flash / OpenCode from 72.4 to 77.5 average; k=5 parallel recursion reaches 80.8, outperforming ReasoningBank, Meta-Harness, and Harness Scaling baselines.

How It Works

1. A task arrives

The executor starts a fresh run from the base harness H₀; prior experience resides only in the improver's library.

2. Reference and propose

The improvement harness executes the base harness as a reference, then proposes invariant and domain-specific candidate edits to the harness components.

3. Probe and compose

Each candidate is executed to gather evidence; a composer reads all candidates and execution evidence, retaining only edits supported by results, and produces the final harness H′.

4. Execute, reflect, and update

The final harness executes and delivers the task output. The improver then reflects on the full improvement record and updates the library of reusable harnesses and lessons for future tasks.

Pros & Cons

Pros

Improves agent scores with the model weights frozen — no fine-tuning required
Consistent gains across all four reported model–harness configurations (Gemini 3.5 Flash and GLM-5 with OpenCode and Codex)
Gains come from an interpretable artifact: markdown harness files rather than opaque weight updates
Outperforms reported baselines (ReasoningBank, Meta-Harness, Harness Scaling) on the WorkBuddy average
The base harness is reset each task, so per-task execution stays anchored while learning lives in the library

Cons

Who It's For

Best for

AI researchers studying agent harnesses, self-improvement, and memory/skill accumulation
Teams with a fixed (possibly closed-source) model who want better agent behavior without fine-tuning
Practitioners using the OpenCode or Codex harnesses with Gemini 3.5 Flash or GLM-5 executors
Anyone reproducing or building on the arXiv paper via the published code and BibTeX

Not ideal for

Teams wanting a turnkey commercial product with support, dashboards, or hosted deployment
Workloads where extra inference calls per task (reference runs, probes) are prohibitively expensive

Use Cases

Improving agent performance on coding, web, and office-style tasks without fine-tuning the underlying model

Research into agent harness design and self-improving agent systems

Building agents that accumulate reusable harness components and lessons across a stream of tasks

Comparing against memory/skill-learning methods like ReasoningBank or meta-harness approaches on agent benchmarks

FAQ

What is Harness²?

Harness² is a method for recursive agent harnessing: it improves an agent by refining the harness around a frozen model (system prompt, guardrails, memory, tools, subagents, skills, scripts, plugins), then updates an improvement library so future harness refinements benefit from past experience.

Does Harness² change the model?

No. The model weights stay frozen. All improvement happens in the editable harness components, so the same executor model scores higher without fine-tuning.

What results does it report?

On the WorkBuddy benchmark, Gemini 3.5 Flash with OpenCode improves from 72.4 (base harness) to 77.5 after one improvement step, and to 80.8 with parallel recursion at k=5, across the Code, Web, and Office domains. Gains are also reported on JobBench and LAB.

Which models and harnesses are supported?

Harness²