Core Argument
Coding agent performance depends not solely on model quality but significantly on the surrounding “harness” — the operational environment including instruction files, bootstrap scripts, validation checks, and test commands. “The difference is not just the model. It is the harness around the model.”
The Problem
Teams typically optimize harnesses manually through trial-and-error adjustments, making it difficult to determine which changes actually improve performance. Meta-Harness (the library) treats the harness itself as an optimization target rather than focusing exclusively on prompt engineering.
Design Philosophy
Meta-Harness uses a filesystem-first approach, storing all candidate workspaces, proposals, validation results, and artifacts on disk. This preserves a complete audit trail for inspection and comparison.
Core Features
- Minimal optimization engine — focused on harness search, not general-purpose optimization
- Filesystem-backed run storage — all state on disk for transparency and debugging
- Write scope enforcement — restricts which paths the agent can edit during optimization
- Environment bootstrap snapshots — captures system state before agent runs
- Explicit candidate outcome classification — structured evaluation of each harness variant
- CLI interface and experiment matrices — scriptable experimentation
Provider Support
- Codex (hosted and local Ollama) — primary validated backend
- Gemini, Pi, OpenCode — implemented experimentally, lack production validation
Getting Started
Installation: uv tool install superagentic-metaharness
Context
Inspired by Stanford AI Lab’s Meta-Harness research paper (Lee et al., arXiv:2603.28052). This is an alpha release prioritizing one validated provider path over premature multi-provider claims.