
Challenges with AI Agents Reliability

“Engineers are becoming sorcerers” | The future of software development with OpenAI's Sherwin Wu
Topic 3 of 14
Published Feb 12, 2026 · Clip 12:43–15:11 (2:29)
Sherwin shares experiences of agent unreliability causing stress, highlighting an OpenAI experiment with a 100% Codex-written codebase without human coding escape hatches. Key issues stem from insufficient context, requiring better documentation like code comments or MD files. Learnings include encoding tribal knowledge to aid agents, with upcoming blog posts on best practices.