Simon Willison summarizes 2026 AI agent developments

Summary

Simon Willison's keynote summarizes the year of large language models (LLMs) in 2026. Coding agents have become reliable for daily use with Claude Opus 4.5 and GPT-5.1. The 'Claw' personal agent category has exploded, with OpenClaw reaching over 100,000 commits. Meanwhile, the engineering community is grappling with new psychological states such as 'Deep Blue' fatigue and 'AI mania.'

Why it matters

The article documents a genuine turning point where AI coding tools have transitioned from novelty to production-grade reliability. This shift is reshaping how developers plan projects and perceive what is possible. The rapid emergence of the agent ecosystem—including personal agents, agent social networks, sandboxing efforts, and 'software factories'—indicates that the 2026 landscape is defined not just by model comparisons, but by agentic workflows and security surfaces.

Highlights

• Coding agents have moved from a 'frequently errors' state to being 'reliable enough for daily use' following model releases in November 2025.
• OpenClaw changed its name four times in less than two months. With over 100,000 commits, it became the most 'vibe-coded' project to exist and spawned an entire 'Claw' software category.
• Willison uses the terms 'Deep Blue' for AI-induced engineering fatigue and 'AI mania' for the pressure to keep agents running continuously. These phrases frame the psychological turbulence of the year.

Critical take

The text reads more like a collection of personal anecdotes than a structural analysis. The 'Pelican on a bicycle' analogy is, in the author's own words, 'the stupidest analogy in the world.' The OpenClaw narrative—featuring sold-out Mac Minis and Meta buying an agent social network in a month—relies on viral spectacle rather than lasting architectural insights. Additionally, the article cuts off mid-sentence in the Dark Factory paragraph, suggesting it is an unpolished keynote transcript rather than a finished critical essay.