Ryan Lopopolo
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Ryan Lopopolo is an engineer at OpenAI who, as of his talk, had spent nine months building software exclusively with AI agents. He is the originator of the "Harness Engineering" framework — a methodology for structuring agent-driven software development around the principle that "humans steer, agents execute."
Core Philosophy: Scarcity Has Shifted
Lopopolo's foundational argument is that the economics of software engineering have fundamentally changed. He claims that "implementation is no longer the scarce resource of software engineering; human time, human attention, and model context window are the true scarce resources." (5:11) This leads directly to his corollary that "code is free" — there is an abundance of code to solve problems (3:35) — and that code should therefore be treated as "a disposable build artifact rather than something to be carefully conserved." (41:44)
The Harness Engineering Framework
Lopopolo advocates for a holistic approach he calls Harness Engineering: how to build software when humans steer and agents execute. (0:17) The harness is not just tooling but the entire system of constraints that shapes agent output. He offers an analogy: "The LLM is analogous to a fuzzy compiler where the harness constraints act as optimization passes that determine what code is acceptable output." (42:38) Critically, he argues this investment is durable: "The harness engineering approach of providing context at the right time is not made obsolete by increases in model capability." (43:29)
First-Party Tool Leverage
Lopopolo advocates using Claude Code as the entry point to the development process, rather than building a custom environment around it. (20:20) His reasoning is that "models are post-trained in the context of the harness in which they are primarily deployed, so depending on first-party harnesses gives additional leverage." (27:03)
Repository Structure as Agent Infrastructure
A major theme in Lopopolo's work is encoding context structurally into the repository so agents don't require manual guidance. He operates a Monorepo with 750 packages in a PNPM workspace isolated by business logic domain (57:20), arguing that "structuring repositories into many small, isolated packages improves agent performance by creating transferable context and scoping changes to directory subtrees." (34:02)
He pairs this with Codebase Uniformity — "making the code as much the same as possible" (58:50) — on the grounds that having one canonical pattern "reduces the model's required attention and makes token output more predictable." (35:28) ESLint custom rules wired into every package in the workspace serve as a mechanical enforcement layer for these patterns. (21:10)
Progressive Disclosure of Context
Rather than front-loading all instructions, Lopopolo advocates Progressive disclosure: "figuring out ways to either defer or just-in-time surface those instructions," because overwhelming the agent upfront is counterproductive. (40:57) He uses GitHub markdown files in the repository as the primary "hub and spoke" mechanism for delivering this context. (47:21)
Skills: Depth Over Breadth
Lopopolo advocates maintaining a small, focused set of Skills — roughly 5 to 10 — and making them progressively better rather than expanding the surface area broadly, "because tooling changes too fast to maintain a wide skill surface." (22:26)
Review Agents and Continuous CI
He uses Review agents — security and reliability agents running continuously on every push and in CI — as a systematic quality gate rather than relying on human review to catch agent mistakes. (12:03, 20:05)
Garbage Collection Day
Lopopolo institutionalized a practice called Garbage collection day: one day per week (Fridays) dedicated to cleanup. (37:54, 1:02:34) This reflects his broader view that human review feedback identifying agent mistakes should trigger systematic fixes — "encoding that context into the repository rather than relying on ongoing manual review." (38:30)
Human Oversight Principles
Despite his enthusiasm for agent autonomy, Lopopolo draws firm lines around human attention. He argues that "every time a human must type 'continue' to an agent is a failure of the harness to provide enough context around what it means to continue to completion." (33:13) At the same time, he insists on mandatory human review of agent plans before execution: "approving agent plans without reading them encodes bad instructions that waste subsequent execution." (41:17)
Points of Disagreement
Lopopolo's stance on Plan Mode is notably skeptical: he states he hasn't "really used plan mode as part of the harness at all," (40:54) which sits in tension with workflows advocated by others who treat plan-then-execute as a central agent pattern. His dismissal of plan mode is paired, however, with a strong insistence that plans — when used — require mandatory human review rather than rubber-stamp approval.