HarnessEngineering
Teams trying to get better results from AI coding agents often reach first for a better model or a cleverer prompt. Harness engineering is a named practice that holds the model and agent fixed, as a black box, and instead improves the two things around it that can actually be controlled: the context the agent is given and the tools it can use, along with the environment curated around them. The argument is that a general-purpose model only contains the visible tip of an organization's knowledge, and the rest, its conventions, quality bar, procedures, exception history, and lines of authority, has to be supplied deliberately or the agent cannot do a particular job well. The practice is collected in Ryan Lopopolo's harness-engineering anthology and field guide, a CC-BY-4.0 body of writing meant to be read both by people and by agents.
General model weights hold only what is publicly common; an organization's actual operating knowledge sits below the waterline. That includes current system state, local vocabulary, the quality bar, standard procedures, the record of past exceptions, and who is allowed to decide what. The practice frames the core task as making this private, changing process data available to a capable worker as retrievable context and usable tools, rather than assuming the model will intuit it.
- Treat the model and coding agent as a fixed black box; improve context and tools, not the weights.
- Encode nonfunctional requirements (reliability, security, compatibility, maintainability, performance, operability, risk posture) and the local decisions about trading them off into retrievable context, worked examples, tools, and executable checks.
- Make the feedback loop cumulative: lessons from accepted work, corrections, and failures become boundaries, examples, and checks that shape later runs, so organizational judgment compounds across agent-maintained artifacts.
Harness engineering is the umbrella discipline around several narrower entries. It is distinct from Harness Handbook, which is a specific method for documenting an existing harness so it can be audited and edited, and from Self-Improving Agent Harness, which is an experiment in letting an agent modify its own harness. Both of those sit inside the broader practice this entry names.
Assess. The material is a curated single-author anthology (Ryan Lopopolo, roughly 900 stars, CC-BY-4.0) with modest adoption, explicitly framed as a field guide and set of intuitions rather than a productized or independently validated method. That places it in Assess with no first-person use. Center position reflects that the transferable value is the framing and vocabulary it supplies rather than a runnable artifact to trial directly.
