Harness Engineering argues context beats model tweaks

Ryan Lopopolo’s Harness Engineering makes a simple argument: if you want better results from coding agents, stop treating model upgrades as the main lever. Keep the model and agent fixed as a black box, and improve the environment around them instead.

Lopopolo defines harness engineering as the practice of shaping the two external inputs that matter most: context and tools. The goal is to build an environment where an agent can recover intent, act on real systems, respect authority, prove what it did, and leave behind artifacts that help the next run.

A core part of that environment is an organization’s nonfunctional requirements — the constraints and quality targets around reliability, security, compatibility, maintainability, performance, operability, risk posture, and polish. The harness also captures local decisions about how those requirements are prioritized and traded off in practice.

Recommended reading

Open models are overtaking AI’s front line

Lopopolo says this systems-level framing comes from 2026's [un]prompted conference, describing it as a way to get “the whole universe of nonfunctional requirements into code.” In his telling, the repository becomes a teaching layer for the agent: requirements and decisions are turned into retrievable context, examples, tools, and executable constraints.

That matters because agent work is iterative. Accepted changes, failures, corrections, and user feedback can all be fed back into the harness so organizational judgment becomes cumulative over time. Rather than relying on general model weights to infer a company’s internal norms, the harness provides the private and fast-changing process data those models typically lack.

Lopopolo frames that missing layer as the bulk of the iceberg below the surface: an organization’s current operational state, local ontology, procedures, exception history, authority relationships, and quality bar. His point is that agents will not reliably infer which of that matters unless it is explicitly made available.

The repository is meant to be used directly with a coding agent. Lopopolo says AGENTS.md routes tasks to the relevant arguments, cases, and proof, while readers who want the broader argument can start with the thesis index or move to the playbooks for practical application.

The repository-authored material is licensed under CC BY 4.0, with attribution and source-material rights detailed in COPYING.md.