Simon Willison

Bio

Simon Willison's foundational contribution is the taxonomy of how professionals actually work with LLMs, distinguishing between "vibe coding" (fast, loose, prompt-driven, no accountability for how code works) and what he initially called "vibe engineering" before settling on "agentic engineering" -- where seasoned professionals accelerate their work with LLMs while staying accountable for production-quality output. This distinction matters because it clarifies that the highest-value use of AI is not replacing expertise but amplifying it.

Operating themes

• Operating thesis: AI tools amplify existing expertise -- the more skills and experience you bring, the better results you get -- and the discipline of working with coding agents is closer to engineering management than to programming.
• Building With Llms
• Prompt Engineering
• Agentic Ai Design

Cards

• November 2025 was the qualitative threshold, coding agents now almost always do what you tell them, Coding agents crossed the "almost always works" threshold in Nov 2025 [Tier A]
• The dark factory: nobody reads the code, gated by a simulated QA swarm, Nobody types or reads code; simulated QA swarm gates correctness [Tier A]
• Hoard a personal repository of things that worked, coding agents will recombine them, Hoard a personal repo of working artifacts; the agent recombines them [Tier A]
• Encode jargon shorthand once, save tokens forever, Encode jargon shorthand once, save tokens forever [Tier B]
• As coding agents become more reliable, review discipline erodes and the failure mode becomes invisible, As coding agents become more reliable, review discipline erodes [Tier A]
• Running MicroPython inside a WebAssembly sandbox gives AI agents safe code execution without host system access risk, MicroPython in a WASM sandbox gives AI agents safe code execution [Tier B]
• Per-employee AI tool spending caps anchored to compensation signal an implicit enterprise ROI model, and are more sustainable than consumption competitions, Per-employee AI tool spending caps anchored to compensation signal an implicit ROI model [Tier C]
• Silent AI model degradation breaks user trust more fundamentally than explicit refusals because users cannot adapt to limitations they cannot detect, Silent AI model degradation breaks user trust more fundamentally than explicit refusals because users cannot adapt to limitations they cannot detect [Tier B]
• Frontier AI agents deploy every available technique to reach their goal, using breadth-first problem decomposition when one approach is blocked, Frontier agents deploy breadth-first problem decomposition, trying every available technique when blocked [Tier B]
• Publish writing while still unhappy with it; the perceived flaws are invisible to readers and the only alternative is a folder of permanent drafts, Publish writing while still unhappy with it; perceived flaws are invisible to readers and the only alternative is a folder of permanent drafts [Tier B]

• Coding agents shift the difficulty floor upward, leaving only the hard problems for human engineers, Coding agents shift the difficulty floor; only hard problems remain for the engineer [Tier B]
• Directing AI agents demands the same skill that defined good engineering. Defining goals clearly and choosing tools well is what engineering always was., Directing agents demands the same skill that made good engineers good [Tier B]

Sources captured

• 2026-04-02, Lenny's Podcast, "Agentic engineering and the November inflection" (raw/podcasts/simon-willison--agentic-engineering-november-inflection--2026-04-02.md)
• 2026-04, heres-how-i-use-llms-to-help-me-write-code.md (operator essay archive)
• 2026-04, vibe-engineering.md (operator essay archive)
• 2026-04, agentic-engineering-patterns-simon-willisons-weblog.md (operator essay archive)
• 2026-09-24, simonwillison.net, "Note on 24th September 2026"
• 2026-09-27, simonwillison.net, "2026 in LLMs (so far)" (keynote at WeAreDevelopers World Congress)