The 70% problem, as framed in Addy Osmani’s agentic-engineering glossary, is the observation that AI coding tools reliably generate a working first draft — structure, boilerplate, happy-path logic, standard patterns — covering roughly 70% of a task almost immediately, while the remaining 30% (the null-instead-of-empty-array edge case, a race condition that only appears under load, an ignored accessibility requirement, a security vulnerability hidden behind plausible-looking code, an awkward integration point) takes longer than the first 70% did and requires engineering judgment that AI can’t reliably supply on its own.
Usage
The source presents the 70% problem as the reason agentic engineering exists as a discipline: if AI could reliably deliver the full 100%, no engineering practice would be needed around it. Recognizing the gap changes how an engineer allocates their time, shifting effort from writing code to reviewing, testing, handling edge cases, and integrating it — described as a fundamentally different, and arguably harder, skill than writing code from scratch. The 70% problem is framed as applying mainly to production-bound code; AI is described as excelling at prototypes, where the 70% is often sufficient. Better prompts, specs, and context are said to push the AI-covered share from 70% toward 80–85%, but the source states that closing the full gap still requires human engineering judgment. It also frames junior developers who only know how to reach the 70% mark as less valuable on a team than senior engineers who can efficiently close the remaining 30%.
Related Terms
The source names this term alongside Human-in-the-loop, Comprehension Debt, Skill Atrophy, and DefinedTerm/ai-assisted-code-review as related glossary/blog entries, without defining any of them.