saved
Harness Engineering: an anthology and field guide
gist
Harness Engineering is Ryan Lopopolo's public anthology, field guide, and agent-context bundle for improving coding-agent output without changing the model. It treats the chosen model and agent as a black box and works the two external levers, context and tools, so the repository can carry an organization's nonfunctional requirements and local tradeoffs. A capable worker should recover intent, operate the real system, respect authority, prove the outcome, and leave the next run better equipped. General weights hold only the tip of that private process-data iceberg; last-mile harness work makes the rest available.
ideas
- Hold the model constant. Improve output by shaping context and tools around a chosen model and coding agent treated as a black box.
- The repository carries nonfunctional requirements. Reliability, security, compatibility, maintainability, performance, operability, risk, and polish should become recoverable context, examples, tools, and executable constraints.
- Judgment can accumulate. Accepted work, corrections, failures, and user responses become later context, boundaries, and checks so organizational coherence compounds.
- Weights are only the tip. Current operational state, local ontology, quality bar, procedures, exception history, and authority sit below the waterline and will not appear in general models.
- Point the agent at the bundle. AGENTS.md routes a coding agent to arguments, cases, and proof; the thesis index and playbooks are the human reading path.
quotes
“It improves the two external levers—context and tools—and curates the environment around them.”
“The worker should be able to recover intent, operate the real system, respect authority, prove the outcome, and leave the next run better equipped.”
“General model weights contain only the visible tip of an organization’s process-data iceberg.”
“Harness engineering is the last-mile work of making it available to a capable worker as context and tools.”