Descriptions:
Ryan Lopopolo, a member of technical staff at OpenAI, delivered a talk at the AI Engineer conference in London on what he calls “harness engineering” — the discipline of building repository infrastructure specifically designed for AI coding agents to operate within. Drawing on nine months of building software exclusively through agents, Lopopolo argues that implementation is no longer the scarce resource in engineering: code is free and infinitely parallelizable, while human time, attention, and model context windows are the real constraints.
The talk goes deep on the practical mechanics. Lopopolo’s team uses OpenAI Codex as the primary entry point into the codebase — agents invoke the dev environment rather than humans wrapping agents inside it. Supporting infrastructure includes custom ESLint rules wired into every PNPM package, a local observability stack with logging and telemetry, a Chrome DevTools integration via a local CLI daemon, and higher-level structural tests that assert things like package privacy, dependency edge correctness, and canonical implementation of async helpers. These guardrails catch the tendency of agents to optimize for local package coherence at the expense of shared utilities, keeping human reviewers from burning attention on low-level drift.
For engineering leaders and senior developers navigating the shift to agent-driven development, this talk provides one of the most concrete operational frameworks available — covering how to restructure repos for agent consumption, how to shift team processes toward delegation and asynchronous review, and why the staff-engineer mental model (directing many parallel contributors) is the right frame for this moment.
📺 Source: AI Engineer · Published April 17, 2026
🏷️ Format: Deep Dive
