The 80% Problem in Agentic Coding – Addy Osmani
4 days ago
- #AI-coding
- #productivity
- #software-development
- Andrej Karpathy and Boris Cherney highlight a shift to 80%+ AI-generated coding, with some developers relying entirely on AI for code.
- AI coding errors have evolved from syntax issues to conceptual failures, including assumption propagation and abstraction bloat.
- Comprehension debt emerges as a hidden cost, where developers rubber-stamp AI-generated code without deep understanding.
- Productivity paradox: AI increases code output but also review times, creating new bottlenecks in development workflows.
- Effective patterns include agent-first drafts, declarative communication, automated verification, and maintaining architectural hygiene.
- Skill atrophy is a risk for heavy AI users, emphasizing the need for deliberate learning and maintaining coding fundamentals.
- The shift from imperative to declarative coding leverages AI's ability to iterate until success criteria are met.
- The 'slopacolypse' risk highlights concerns about signal-to-noise ratio as AI-generated content proliferates.
- Developers are splitting into those who enjoy coding as a craft and those who focus on building, with tooling optimizing for the latter.
- AI amplifies both good and bad development practices, making robust architecture and thorough testing more critical than ever.