Simon Willison Blog Digest — January 2026
Consolidated insights from 20 posts (Jan 22 – Feb 1, 2026) extracted from the ops-core buffer before 30-day expiry.
Coding Agents & AI-Assisted Programming
Scale vs. Skill Debate
Two posts presented contrasting approaches to agent-driven development. Wilson Lin at Cursor used thousands of parallel agents to build FastRender, a web browser in Rust. Days later, embedding-shapes built a functioning HTML/CSS browser (20,000 lines of Rust, zero dependencies) with a single agent in 3 days. The takeaway: a skilled engineer with one agent can match sophisticated multi-agent systems.
Tips for Agent-Written Tests
Python’s strong pytest ecosystem gives agents excellent training data. Best practice: show agents existing codebases with good patterns rather than writing instructions. “Clone datasette/datasette-enrichments and imitate patterns” outperforms detailed test specs. Use pytest.mark.parametrize and fixtures to reduce duplication.
Design for Agents, Not Humans
Steve Yegge described the “desire paths” approach: Beads evolved 100+ CLI subcommands by making agent hallucinations real. The complex CLI isn’t for humans — it’s designed so agents succeed at what they naturally try. Iteratively implemented whatever agents attempted until “nearly every guess is now correct.”
Five Levels of AI Autonomy
Dan Shapiro’s taxonomy: (1) Spicy autocomplete, (2) Coding intern, (3) Junior dev, (4) The developer, (5) Engineering team, (6) Dark factory. At level 5, nobody reviews AI code — systems prove correctness via testing and simulation. Small teams (<5 people) achieving “nearly unbelievable” results at this level.
Economic Threshold Shift
Chris Ashworth (QLab creator) used Claude Code to build a custom lighting design app serving an audience of 3. Previously never economically viable. Key caveat: “It can take a bad programmer and make them faster at making bad programs” — understanding code quality remains essential.
Process Skepticism
Jenny Wen (Anthropic Design Lead) argues AI makes prototyping cheap enough to explore multiple wrong directions without catastrophic waste. “In a world where anyone can make anything, what matters is your ability to choose and curate.”
Browser Sandboxing & Security
Paul Kinlan (Google) made the case that 30 years of browser sandbox engineering positions browsers as the ideal AI agent runtime. Three pillars: File System Access API, CSP with sandbox="allow-scripts allow-same-origin", and WebAssembly in Workers. The Co-do demo showed agent-driven file editing without multi-GB local containers. Meanwhile, ChatGPT’s Code Interpreter quietly expanded to support Bash, Node.js, and 10+ languages with pip/npm via custom proxy — but container.download introduces potential data exfiltration vectors.
AI Capabilities & Cost Deflation
Andrej Karpathy highlighted that training GPT-2 (2019: 73 in 3 hours on a single 8xH100 node — 600X cost reduction, roughly 2.5X cheaper every year. Kimi K2.5 demonstrated multimodal agentic intelligence: a 1T-parameter model with self-directed swarms of up to 100 sub-agents making 1,500 tool calls, achieving 4.5X speedup over single-agent operation.
AI Agents in the Wild
Moltbook emerged as an AI social network where OpenClaw agents interact via skills (markdown instructions + scripts). Agents self-install skills via curl and execute heartbeat instructions from the internet every 4 hours. Simon flagged this as “most likely to result in a Challenger disaster” due to prompt injection risk — the lethal combination of autonomous agents, self-installing skills, and network access.
Tooling & Infrastructure
- Datasette 1.0a24 added multipart file upload support and switched to uv for development
- exe.dev hosts 25 VMs sharing 2 CPUs/8GB RAM for $20/month, routing SSH traffic via public key identification (SSH has no Host header equivalent)
- Simon demonstrated adding dynamic features to aggressively cached static sites using localStorage-based conditional rendering, built entirely via Claude Code from an iPhone
Recurring Patterns
- Agents learn from patterns, not instructions — show existing codebases, not specs
- The “wrong direction” is now cheap — AI collapses cost of building the wrong thing, enabling exploration over planning
- Browser as universal runtime — 30 years of security work positioning browsers as ideal agent sandbox
- Quality requires human expertise — automation amplifies skill level; understanding remains essential
- Undocumented capabilities — frontier models evolving faster than documentation
Related
References
- Simon Willison’s Weblog
- Original posts (Jan 22 – Feb 1, 2026)