An AI state of the union: We’ve passed the inflection point, dark factories are coming, and automation timelines

Simon Willison, a veteran software engineer and co-creator of Django who writes the widely read simonwillison.net blog, discusses how coding agents crossed a reliability threshold in late 2025 and what that means for software work. The conversation covers agentic engineering practices, the 'dark factory' idea, security risks like prompt injection, and the personal toll of working with AI agents.

01Key learnings

• Coding agents crossed a reliability threshold around November 2025, so they now usually do what you ask, which changed how seriously engineers can take them.
• Writing code is now cheap, so the bottlenecks move to ideation, prototyping, testing, and process; build several quick prototypes of an idea and compare them.
• Use test-driven development with agents: have them write tests, run them, and watch them fail first (Red/Green TDD) to keep quality and avoid breaking old features.
• Keep a personal backlog of tried techniques and tools in GitHub (tools and research repos) so agents can reuse prior working code as context for new problems.
• Start new projects from a thin, well-structured template, since agents closely follow existing patterns in a codebase.
• Vibe coding is fine for personal throwaway tools, but code that could harm others needs professional review; reserve 'agentic engineering' for production work.
• Prompt injection is effectively unsolved; limit an agent's blast radius by cutting one leg of the lethal trifecta (private data, untrusted input, exfiltration) and avoid running agents with sensitive data.
• Agentic work is mentally exhausting; expect a limit on how many agents you can supervise, and watch for burnout and gambling-like overuse.

02Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.