Addy Osmani is an Irish software-engineering leader, author, and open-source developer whose career spans Chrome developer tools, web performance, and AI-assisted software development. During more than 14 years at Google, he led developer-experience work across Chrome and Gemini before becoming a director at Google Cloud AI. He now builds independent educational and open-source projects addressing a central challenge of agentic engineering: keeping humans accountable for software increasingly produced by machines.
Osmani built a browser called XWebs as a teenager in Ireland. At Google, his Chrome work encompassed Chrome DevTools, Lighthouse, and Core Web Vitals; he was an early member of the team behind Lighthouse, which helps developers audit website performance, accessibility, and quality.
His open-source project Critical extracts and inlines the CSS needed for initial page rendering. In 2019, he explained Chrome’s native image lazy-loading, which delays loading offscreen images. His openly licensed Backbone.js book and books including Learning JavaScript Design Patterns and Leading Effective Engineering Teams extend that practical focus into engineering education.
Osmani moved from Chrome to Google Cloud AI in December 2024, leading developer experience and technical evangelism across work involving Gemini and Google’s agent-development platform. After leaving Google in 2026, he created Agent Engineer, a free course on building and shipping AI agents, and agent-skills, a repository of production-oriented guidance for coding agents.
What accountable agentic engineering requires
- Agent harnesses and engineering loops: Useful coding agents need context, tools, filesystem access, version control, evaluation, and recovery mechanisms. Together, these systems can investigate problems, implement changes, run tests, and produce evidence while engineers determine what deserves attention.
- Human ownership of the outer loop: Agents can generate diffs, logs, tests, and screenshots; people must determine whether that evidence is sufficient, accept production risk, and remain answerable when software fails. Osmani argues that human judgment shifts toward approval and accountability as execution becomes automated.
- Cognitive debt and cognitive surrender: Teams accumulate cognitive debt when generated code grows faster than their understanding. Cognitive surrender occurs when engineers adopt an agent’s conclusion before forming an independent judgment. Passing tests cannot compensate for losing the ability to explain a system or challenge a flawed result.
- Intent debt: Undocumented goals, constraints, and design rationale cannot reliably be recovered from code alone. Building on Margaret-Anne Storey’s existing framework, Osmani treats durable specifications and decision records as essential infrastructure for agent-assisted development.
- The orchestration tax: Parallel agents do not multiply human attention. Every merge conflict, ambiguous result, and production decision draws on limited review capacity, making agent concurrency a supervision problem as much as a throughput opportunity.
His AI Engineer World’s Fair keynote places engineering responsibility at the boundary between automated execution and human judgment: deciding what should exist, verifying what was built, and owning what ships.
