Google Open-Sources Harness²: Dual-Layer RSI Framework Boosts Agent Performance Without Weight Updates

Why Not Modify Model Weights?

Agent deployment in real-world scenarios faces three fatal problems with fine-tuning: prohibitive compute and annotation costs, catastrophic forgetting of prior capabilities, and inability to modify closed-source models like Gemini or Claude. Harness² provides a non-invasive upgrade path by optimizing the agent's execution framework — the "harness" surrounding the language model — rather than its parameters.

What Is Harness? The Eight Editable Components

Harness² concretizes the agent's infrastructure into eight editable components that act as the agent's "hands, eyes, memory, and toolbox":

1. System Prompt: Defines agent role and base rules.

2. Guardrails: Sets safety boundaries to prevent dangerous operations.

3. Memory: Stores past experiences and context.

4. Tool Guidance: Specifies each tool's usage and pitfalls.

5. Sub-agents: Decomposes complex tasks for specialized smaller models.

6. Skills: Encodes fixed operational procedures for specific scenarios.

7. Scripts: Auto-executed code snippets.

8. Plugins: Interfaces to external services.

The framework supports OpenCode and Codex executors, enabling direct execution and verification of complex code logic. Crucially, Harness² optimizes these components, not model weights, letting the model discover and fix its own tool misuse or prompt deficiencies.

Deep Dive: Dual-Layer Recursive Optimization Architecture

Harness²'s core innovation is its dual-layer recursive optimization , addressing both immediate tasks and long-term capability accumulation.

First Layer: Task-Level Optimization (PPC Mechanism)

When facing a novel task where the agent fails, Harness² employs a three-step PPC (Propose–Probe–Compose) mechanism:

Propose: Analyze the execution trajectory, identify pain points, and generate multiple proposals to modify the eight components.

Probe: Test each proposal by running the agent with the modified components, comparing execution outcomes to identify the best fix.

Compose: Fuse the top-performing modifications into a "ultimate battle suit" for the current task.

Task-level recursion operates in two forms: (a) parallel recursion — simultaneously exploring multiple candidates, each independently composing the next-generation framework; (b) sequential recursion — each step feeds the improved framework and accumulated execution evidence into the next step.

Second Layer: Meta-Level Optimization (Experience Distillation)

After each successful task-level optimization, Harness² extracts core logic and distills experience — e.g., "when processing tabular data, cleaning nulls with a script first improves success rate." These insights are stored with supporting execution trajectories as evidence. When encountering new tasks, the meta-level loop retrieves relevant past patterns to guide Harness editing, enabling transfer learning akin to human analogical reasoning.

Benchmark Results: Single-Step Gains on Professional Benchmarks

With model weights frozen and no ground-truth labels used, Harness² was evaluated on GLM-5 and Gemini 3.5 Flash across challenging benchmarks:

WorkBuddy-Bench (complex office workflow collaboration): up to +11.5 points .

JobBench (realistic professional tasks): steady +4.5 points .

LAB (hardcore benchmark): accuracy surged +11.8 percentage points .

These gains come from only single-step optimization. The paper shows performance further improves as the recursive budget k (allowed self-iteration rounds) increases.

Industry Resonance: Alignment with LangChain's Trace & Eval Paradigm

LangChain core member Sydney Runkle highlighted that Harness²'s design aligns with the frontier agent engineering principles of Trace (trajectory evidence) and Eval (verification of improvements) :

Task-level loops analyze traces to propose improvements; meta-level loops learn which improvement patterns are truly effective in evaluation. This is the essential path to reliable production agents!

Harness² effectively provides a fully automated, self-consistent "agent doctor" atop frameworks like LangChain, eliminating manual log inspection and prompt tuning.

References

GitHub: https://github.com/google-research/harness2 arXiv: https://arxiv.org/abs/2610.14224 Project page: https://harness2.github.io/ Sydney Runkle's X post:

https://x.com/sydneyrunkle/status/2112043827553005651