Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI
TL;DR
OpenAI engineer Ryan Lopopolo explains how AI agents have made code generation abundant and free, requiring a fundamental shift from writing implementations to designing systems, documentation, and guardrails that enable agents to execute software engineering tasks autonomously.
💰 The Abundance Paradigm 3 insights
GPT-5.2 marked the "magic moment" for full agent autonomy
Lopopolo identifies GPT-5.2 as the capability inflection point where models became isomorphic to human software engineers in producing production-quality code.
Human engineers now command thousands of agents
Individual engineers can deploy five to five thousand agent-workers in parallel, constrained only by GPU capacity and token budgets exceeding one billion output tokens daily.
Parallel execution eliminates priority backlogs
Previously abandoned P3 tasks can now be attempted immediately across multiple parallel trajectories, with successful solutions merged and failures discarded.
🎯 Steering Through Documentation 3 insights
Documentation displaces code as the primary artifact
The valuable output is no longer the implementation but the prompts, ADRs, and guardrails that guide agents, as code has become free to produce, refactor, and delete.
Encode non-functional requirements explicitly
Teams must document quality standards for security and reliability as lint rules and review agent prompts to prevent agents from generating maintenance-burdened "slop."
Durable expertise scales through written standards
Documenting QA plans and coding standards once allows every agent trajectory to inherit the team's collective expertise without synchronous human review.
🤖 Operationalizing Agent Teams 3 insights
Deploy automated reviewer agents in CI
Security and reliability agents continuously check proposed patches for patterns like missing timeouts, retries, or unsafe interfaces before human involvement.
Engineer codebases for context efficiency
Adapt repositories to model constraints by enforcing file size limits and writing error messages that include specific remediation prompts to maximize limited context windows.
Compose stackable agent skills
Create reusable "skills" that encode best practices—such as prompt engineering—which agents invoke to maintain consistency and build compound leverage.
Bottom Line
Stop writing implementation code and instead build systems of documentation, automated guardrails, and agent skills that enable AI agents to execute the full software engineering lifecycle while humans focus exclusively on high-level steering.
More from AI Engineer
View all
How to talk to statues — Joe Reeve, ElevenLabs
Joe Reeve from ElevenLabs discusses building a viral AI app that lets users talk to statues via phone calls, exploring how vibe coding with existing APIs enables rapid prototyping, the unique challenges of voice interface design, and the cultural implications of giving physical objects AI-generated voices.
How I deleted 95% of my agent skills and got better results — Nick Nisi, WorkOS
Nick Nisi from WorkOS explains how deleting 95% of his AI agent's skills improved accuracy from 77% to 97%, detailing his 'Case' harness system that uses state machines and cryptographic proof to enforce accountability rather than relying on instructions.
Frontier AI at Home — Alex Cheema, EXO Labs
Alex Cheema from EXO Labs argues that AI should function as a local 'exocortex' rather than rented cloud infrastructure, detailing why inference optimization (not training) is the key bottleneck and how exponential improvements in 'intelligence per joule' will make consumer-grade frontier AI feasible within years.
`What the Best Agents Share` — Mardu Swanepoel, Flinn AI
Mardu Swanepoel from Flinn AI analyzes four design patterns shared by top AI agents—focus modes, transparent execution, personalization, and reversibility—to demonstrate how constraining scope, building trust, and reducing downside risk creates more effective human-agent collaboration.