meta-harness

Harness code optimization uses native Pareto search loops to automatically tune memory systems, retrieval policies, context assembly, and prompt templates around a fixed base model, finding configurations that deliver higher quality at lower cost without requiring model weight retraining.

Pareto frontier search optimizes memory systems, retrieval policies, context assembly logic, prompt templates, and summarization scaffolds within the harness code, systematically improving system quality and reducing operational cost across repeated task episodes.

You execute evolutionary optimization loops natively using built-in Agent, Workflow, and loop tools to propose candidate harness variants, score them against a cheap deterministic evaluator, and maintain a Pareto frontier of quality versus cost without relying on external Python harnesses.

Before running automated Pareto optimization, you must establish a cheap deterministic evaluator, a held-out evaluation corpus, a defined candidate interface boundary, and a proposer prior to successfully execute the native optimization loops over your harness code.

A deterministic evaluator prevents data leakage and frozen-replay defects by enforcing strict guardrails during candidate scoring, ensuring harness variants are graded accurately on quality and cost metrics against the held-out evaluation corpus.

Meta-harness effectively reduces operational costs for a fixed-model agent by applying native Pareto frontier optimization to the surrounding harness code, discovering configurations that achieve better performance at lower cost compared to manual tuning methods.