Harness Engineering Theses

These twelve theses explain how context, tools, and the surrounding environment shape the work a fixed model and coding agent can complete. Each page develops the argument through Ryan Lopopolo's articles, talks, interviews, public posts, implementation evidence, and related work.

Start here

Read Harness Engineering for an introduction to the practice. Once a target-local decision is unresolved, use this index to choose the one thesis that can change it; choose a procedure from the playbook index when the task calls for an application.

Theses

Hold the worker constant

Treat the chosen model and coding agent as a black box during one adoption epoch. Requalify the environment, human priors, ambition, and inner-loop latency when that worker changes; preserve useful learned semantics across host migrations.

Deploy into the private process-data iceberg

The large applied-AI opportunity is institutional work that depends on private, changing process data. Organizations cannot presume that this data will enter general model weights or that agents will reliably intuit which parts matter. Keep current records, organizational context, and repository knowledge with authoritative owners, then compose them into retrievable context and bounded capability. Goal-specific context repositories can project those sources for a parent worker without replacing them. Code can remain the worker’s internal action language for users who consume domain outcomes without reviewing the implementation.

Give one agent the whole job

Humans provide direction, judgment, and consequential authority. Agents recover the outcome behind a sparse prompt. One primary trajectory owns decomposition, execution, integration, proof, and safe lifecycle closure.

Route context just in time

Keep a large navigable knowledge store and a small active working set. Root guidance routes across authoritative sources; runbooks preserve workflows; skills teach an approach; checks surface constraints when they become relevant. A context-curator sidecar can stage observed gaps, reconcile them with their owners, and publish a reviewed projection at an atomic revision.

Make capabilities legible and operable

Tools are the second external lever. A capability becomes usable through the complete loop of discovery, selection, invocation, interpretation, repair, and real-system verification.

Make the repository teach the agent

Code the agent reads becomes prompt material. The repository makes local nonfunctional requirements recoverable through canonical owners, repeated structures, completed migrations, examples, types, and executable constraints. The agent can infer what good looks like and transfer that context across the repository.

Maximize autonomy inside explicit authority

Keep capability and authority as separate contracts. Let the worker act broadly in reversible environments; make consequential grants scoped, revocable, auditable, and recoverable.

Prove the outcome in the real environment

Match evidence to the user or operational claim. A green check proves only its assertion; browser journeys, corpora, traces, canaries, and deployed health close different loops.

Turn feedback into infrastructure

Trajectory evidence, MLD, review, and incidents can expose missing harness capabilities. Stable lessons move into their earliest durable owner. Mechanically enforceable principles can drive cleanup across a discoverable population, reducing stale precedent for later work.

Preserve coherence and own lifetime risk

Abundant implementation makes future coherence and lifetime ownership scarce. A feedback-controlled environment carries lived judgment into later work while preserving behavioral contracts, migration integrity, dependency evidence, release identity, and proportional controls.

Run known work as a continuous loop

Settled outcomes with signal, proof, and authority can enter repository-owned loops; unresolved product invention remains foreground work. Durable state, feedback, and retirement conditions keep each run aligned with the current world.

Optimize for measured effectiveness

Tokens, lines, agents, and checks are inputs. Optimize useful outcomes per unit of scarce human attention, with acceptable latency, risk, and maintenance.

Applications

Improve one harnessed job

Observe one representative trajectory, locate its earliest failed handoff, make the smallest authorized and reversible intervention at the owning boundary, and use a fresh rerun to decide whether to retain, revise, or remove it. Test without the intervention when its added value remains unclear.

Review a repository

Follow representative jobs from request through delivery, then locate missing context, capability, ownership, proof, authority, and feedback at their owning boundaries.

Evaluate the harness

Isolate one harness intervention under a fixed worker condition. Keep target truth with its owner, use implementation-neutral contract rubrics, and measure accepted outcomes, proof, architecture, and trajectory cost separately.

Influences and alternate framings

matklad’s architecture writing, Artichoke’s living architecture, “Parse, don’t validate,” and Strangler Fig are earlier influences. Birgitta Böckeler’s Fowler- site article appeared after Ryan’s February 11, 2026 essay and is a later alternate framing.

Evidence

The source library indexes essays, interviews, talks, cases, the public-post corpus, influences, and alternate framings.