The PodKnowledge|← All Episodes
Based on Lenny's Podcast data
Lenny's Knowledge Sketch

95% of My Code,
I Don't Type

Simon Willison
Co-creator of Django · Creator of Datasette · Coined "prompt injection", popularized "AI slop" and "agentic engineering" · 25+ years shipping software
APR 2 2026
The Inflection

The November Threshold

NOVCODING AGENT RELIABILITY
"In November we had what I call the inflection point where GPT 5.1 and Claude Opus 4.5 crossed a threshold. Almost all of the time it does what you told it to do."
  • Before November: agents mostly worked, but you had to pay very close attention
  • After November: engineers who tinkered over the holidays got the "oh wow" moment
  • The reverberations are still shaking software; a bellwether for other information workers
  • Simon now writes much of his code from his phone, walking the dog
Framework

Agentic Engineering,
By the Numbers

SIMON, ALONE4 AGENTS PARALLELWIPED OUT BY 11AM
95%
of Simon's code, not typed by him
10K
lines of code, in a single day
4
agents Simon runs in parallel
  • Every inch of his 25 years of experience gets used steering these agents
  • The best mode is often "unsafe" — letting the agent run without nagging you at every step
  • Coding takes way less time; the bottlenecks have moved to everything downstream
  • By 11 AM Simon is wiped — even without reviewing every diff, there's a cognitive limit
Simon's prediction, refactoredLenny asked when 50% of engineers will have AI write 100% of their code. Simon refactored to 95%, and said it could happen by end of this year in the US, slower in Europe.
The Frontier

The Dark Factory Pattern

  • Today's default: tell the agent what to build, then read every line of code it produces
  • Vibe coding: the opposite — throw it to the wind and see what happened
  • Dark factory: apply professional practices and quality expectations to code you are NOT reviewing directly
  • Origin: the factory-automation term for a factory so automated no humans are on the floor. Lights off.
StrongDM's experiments

Simon points to StrongDM as one of the companies pushing this pattern in the real world, publishing interesting discoveries about what works and what doesn't.

Why it matters

Simon does not just want cheap code. He wants better code — fewer bugs, more features, higher quality — produced with less human bottleneck. The dark factory is the version of that we haven't figured out yet.

Playbook

How Simon Actually Works

  • Use "unsafe" mode where the agent runs without asking permission at every step
  • Grow a research repo. Point new tasks at it: "check out SimonW/research, use the Web Assembly and Rust examples"
  • Coding agents for research beat write-a-report-agents; they actually run the code and plot the graph
  • Ambition dial-up: this year Simon's resolution was to take on more, not less
Writing the book, chapter by chapterSimon is publishing a book on agentic engineering one chapter at a time on his blog — no editor, no publisher, no pressure. He writes when he wants to write.
Contrarian

Simon's Uncomfortable Takes

✗"I don't use this stuff, the code is bad"INSTEAD →✓ Not justifiable anymore. The code is good this year. That excuse held last year; it doesn't hold now.
✗Newer engineers benefit more from AIINSTEAD →✓ Simon spends every ounce of his 25 years of experience steering these agents well. Depth compounds, not decays.
✗You'll spend less time thinkingINSTEAD →✓ Wiped out by 11 AM. Four parallel agents cost you attention even when the code is being written for you.
✗The bottleneck is still coding speedINSTEAD →✓ The bottleneck moved. Now it's specs, review, deployment, and every other process that used to wait for the code to be done.
Based on Simon Willison's episode on Lenny's Podcast. All ideas on this page are from the episode.Watch on YouTubeFollow @simonw on X
GO DEEPER IN THE EPISODE
02:40 The November 2025 inflection point08:01 What’s possible now with AI coding10:42 Vibe coding vs. agentic engineering13:57 The dark-factory pattern20:41 Where bottlenecks have shifted23:36 Where human brains will continue to be valuable25:32 Defending of software engineers29:12 Why experienced engineers get better results30:48 Advice for avoiding the permanent underclass33:52 Leaning into AI to amplify your skills35:12 Why Simon says he’s working harder than ever37:23 The market for pre-2022 human-written code40:01 Prediction: 50% of engineers writing 95% AI code by the end of 202644:34 The impact of cheap code48:27 Simon’s AI stack54:08 Using AI for research55:12 The pelican-riding-a-bicycle benchmark59:01 The inherent ridiculousness of AI1:00:52 Hoarding things you know how to do1:08:21 Red/green TDD pattern for better AI code1:14:43 Starting projects with good templates1:16:31 The lethal trifecta and prompt injection1:21:53 Why 97% effectiveness is a failing grade1:25:19 The normalization of deviance1:28:32 OpenClaw: the security nightmare everyone is looking past1:34:22 What’s next for Simon1:36:47 Zero-deliverable consulting1:38:05 Good news about Kakapo parrots
Every sketch here began as a transcript.

Send us yours and we’ll draw it. $1 a transcript.

Sketch mine →
𝕏︎ X / Twitterin LinkedIn📸 Instagram🔗 Copy link
0:00