Reading up on Simon Willison

Hacker News commenters analyze Simon Willison's experience with Claude Fable's excessive proactiveness in fixing a trivial CSS bug, sparking debate on human agency, learning, and the proper use of AI coding agents.

Claude Fable 5 demonstrably solves difficult coding problems—like compiling CPython to WASM—that stumped earlier models, though evaluation remains subjective and vibe-based.

Simon Willison introduces a new plugin hosting sandboxed HTML+JavaScript apps inside Datasette, enabling custom read/write SQL interfaces within tight iframe constraints.

Simon Willison's blog post documents rapid LLM improvements over six months, using the "pelican riding a bicycle" test as a consistent benchmark to track progress.

The Antfly team shows how coding agents using TLA+ can formally model distributed systems to cheaply find and verify race conditions and other bugs.

The post collects expert perspectives on AI-assisted development, emphasizing verification over generation, harness engineering, function structure, and the cultural divide between software thinking and human experience.

Excessive reliance on AI coding tools creates comprehension debt—a growing gap between code volume and human understanding—that erodes system knowledge and eventually must be paid with interest.

Coding agents can improve code quality by handling refactoring, enabling exploratory prototyping, and compounding small quality improvements over time.

Developers should design APIs optimised for AI consumption, exposing concise documentation via endpoints and using non-destructive candidate writes for safe AI actions.

The Hacker News discussion argues that while AI makes generating code cheap, writing good software remains expensive, and code is a liability rather than an asset.

Simon Willison describes creating a custom macOS presentation app using AI-assisted vibe coding in Swift, which sequences URLs as slides and includes a phone-based remote control via Tailscale.

Simon Willison's guide outlines patterns and principles for effectively using coding agents like Claude Code and OpenAI Codex to improve code quality and developer productivity.

Test-first development (red/green TDD) significantly improves coding agent reliability by ensuring tests fail before implementation, reducing unnecessary or broken code.

Coding agents make writing code nearly free, forcing engineers to reorient habits from cost-saving to quality assurance like testing, documentation, and maintainability.

Nate Meyvis argues that cognitive debt in AI-generated codebases is not inherently worse than in traditional codebases, and AI can help mitigate it.

The article argues that the security risks from AI agent frameworks like OpenClaw are inherent to the architecture of autonomous tool use and marketplaces, not just a single platform.

Simon Willison introduces Showboat and Rodney, CLI tools for coding agents to create Markdown demos and browser automation scripts to verify their work.

StrongDM's AI-driven 'software factory' replaces human code writing and review with agent teams, using scenario holdouts for validation, but faces skepticism around cost, security, and validation limits.

The article argues for the term 'agentic engineering' to describe disciplined, human-oversight-driven AI-assisted development, contrasting it with the reckless 'vibe coding' approach.

Hacker News commenters share varied approaches to automating coding work, from using AI agents like Claude Code and Cursor for TDD and side projects to codifying processes into scripts, while debates emerge on token efficiency, code quality risks, and the irreplaceable value of manual review.

ChatGPT's code execution containers gained the ability to run bash, pip/npm installs, file downloads, and ten new programming languages, a major undocumented upgrade.

The browser's 30-year-old sandbox model can be adapted to safely run AI agents locally, though gaps remain in network isolation and malicious file creation.

The browser's sandbox, tested by decades of web abuse, provides a secure environment for AI agents using the File System Access API.

Autonomous AI coding agents become reliable when tasks have clear, testable success criteria, shifting human oversight from implementation to specification and validation.

Hacker News commenters share mixed first impressions of Claude Cowork, with enthusiastic use cases alongside concerns about data privacy and accessibility for non-technical users.

Simon Willison predicts LLMs will write undeniably good code in 2026, sandboxing will be solved, and a coding agent security disaster is imminent.

Notion AI allowed data exfiltration via indirect prompt injection in March 2025, where edits were saved before user approval, enabling attackers to steal sensitive document contents.

Simon Willison ranked as the most popular blogger on Hacker News for the third consecutive year in 2025, based on analysis of domain score data.

Software engineers must prove code works through manual and automated testing, shifting accountability away from AI agents to the human submitting the PR.

Simon Willison shares patterns for building useful single-file HTML tools with no build step, using LLMs to generate them.

Simon Willison shares patterns for building single-file HTML tools without build steps, emphasizing simplicity and LLM-assisted development.