Lenny's PodcastSimon Willison: Why He No Longer Types 95% of His Code
What happens when 95% of code is written by AI agents on autopilot; dark factory engineering, four parallel agents, and a Challenger-style prompt injection.
CHAPTERS
- 0:00 – 1:27
AI coding agents arrive: 10,000 lines a day and the agent-loop question
Simon opens with how coding agents changed the workflow from “generate code” to “generate, run, test, iterate,” enabling massive output and parallel work. He raises the bigger question of which other knowledge-work domains will soon be susceptible to similar agent loops.
- •Coding agents automate the run/test loop, not just code generation
- •Engineers can now produce enormous amounts of working code quickly
- •Parallelizing work across multiple agents is becoming normal
- •Open question: which other knowledge fields will adopt agentic loops next
- 1:27 – 3:31
Who is Simon Willison, and why this moment matters for software
Lenny introduces Simon’s background (Django, Datasette, prompt injection) and frames him as a practitioner tracking the shift from “old” software building to the new AI-mediated approach. The conversation is positioned as a practical state-of-the-union on how work is changing.
- •Simon’s long track record as a builder and educator (Django, Datasette)
- •Coined/popularized key AI terms and risks (prompt injection, AI slop)
- •Focus on real-world practice over theory
- •Set up: AI’s impact on building software and professional work
- 3:31 – 8:49
The November 2025 inflection point: reasoning models + coding-first training
Simon explains why 2025 became “code is the application” for leading labs, culminating in a threshold-crossing moment in November. Incremental model improvements (reasoning + RL for coding) compounded until agents became reliably useful, changing engineering overnight.
- •Labs prioritized coding because it drove willingness to pay and clear ROI
- •Reasoning models (late 2024 onward) dramatically improved debugging and code planning
- •November releases crossed a reliability threshold: ‘mostly works’ → ‘almost always works’
- •Holiday experimentation triggered a widespread “it actually works now” realization
- 8:49 – 12:22
Vibe coding vs. agentic engineering: where it’s safe, and where it’s not
Simon distinguishes hands-off “vibe coding” (not reading or understanding code) from professional “agentic engineering” (using agents with rigorous review and practices). He argues vibe coding is great for personal prototypes, but dangerous when others can be harmed.
- •Karpathy’s vibe coding: build without reading code; iterate based on behavior
- •Democratizes building small tools and prototypes for non-programmers
- •Responsibility boundary: personal use vs. production code for others
- •Preferred professional framing: “agentic engineering” emphasizes tool-running agents
- 12:22 – 14:54
Dark factories: ‘no one writes code’… then ‘no one reads code’
The discussion moves to the “dark-factory” pattern—highly automated software creation where humans set direction but don’t type (and eventually don’t read) code. Simon details how some companies experiment with rules like “no typing code” and even “no reading code,” pushing quality assurance into new forms.
- •Policy shift: code typed by humans becomes optional or forbidden
- •Next frontier: shipping reliable software without humans reading the implementation
- •Humans still define goals, constraints, and acceptance criteria
- •The challenge becomes professional-grade verification without traditional review
- 14:54 – 21:27
Simulated QA swarms and synthetic integrations: how dark factories test
Simon shares StrongDM’s approach: a constant swarm of agent “users” testing workflows and edge cases, plus simulated versions of Slack/Jira/Okta to avoid rate limits. This reframes QA as large-scale, always-on behavioral testing at high token cost but high coverage.
- •Agent testers simulate end users continuously (24/7)
- •Testing at scale can be expensive in tokens but robust in coverage
- •Building “fake Slack/Jira/Okta” via API docs enables unlimited simulation
- •Demonstrates creative validation methods when code isn’t manually reviewed
- 21:27 – 26:17
Bottlenecks move upstream: prototypes are free, proof is hard
With implementation speed exploding, the slow parts shift to specification, evaluation, and deciding what’s worth building. Simon describes using AI to prototype multiple options quickly, but notes that real usability testing and real-world validation remain difficult to replace.
- •Coding acceleration shifts bottlenecks to ideation, evaluation, and validation
- •Rapidly prototyping 3+ variants becomes practical and powerful
- •AI is strong at brainstorming the obvious, weaker at proving what’s best
- •Human usability testing still beats simulated “AI user” testing for credibility
- 26:17 – 29:29
Why senior engineers still matter: AI as an amplifier (and why it’s exhausting)
Simon argues AI doesn’t eliminate engineering skill—it amplifies it, making experience more valuable for directing agents effectively. He also highlights a new constraint: human cognition and attention become the limiting factor, causing intense fatigue even as output rises.
- •Expertise improves prompting, scoping, and diagnosing what’s feasible
- •Agents enable parallel work, but managing it can burn people out
- •New personal skill: finding sustainable limits and avoiding “always-on” pressure
- •Past intuition about effort/cost estimates becomes unreliable in the agent era
- 29:29 – 31:06
Career impacts: interns ramp faster, mid-level engineers face the squeeze
Drawing on industry observations, Simon notes AI helps both very senior engineers and true beginners (onboarding and productivity). The most vulnerable group may be mid-level engineers who lack deep expertise to amplify but no longer benefit from beginner boosts.
- •Interns become useful faster due to AI-assisted onboarding
- •Senior engineers get disproportionate leverage from their experience
- •Mid-level engineers may be most at risk of displacement or stagnation
- •AI reshapes career ladders and expectations across organizations
- 31:06 – 45:01
Avoiding the ‘permanent underclass’: agency, ambition, and mindful learning
Simon’s advice emphasizes leaning into the tools while deliberately preventing skill atrophy. He frames “agency” as the durable advantage—choosing problems, being ambitious, and using AI to shrink learning curves across domains.
- •Lean in: use AI to learn, expand ambition, and attempt bigger projects
- •Be mindful of skill atrophy; choose when to rely on AI vs. practice yourself
- •Agency as a core human advantage: deciding what to do next and why
- •Practical examples: new languages (AppleScript), even skill-building like cooking
- 45:01 – 48:20
Cheap code changes everything: productivity, prototyping, and lost signals
They unpack the ‘code is cheap’ reality: the time-cost center of software has shifted dramatically, changing team processes and individual workflows. Simon notes a new quality problem: documentation/tests are no longer strong signals if they can be generated instantly without real usage.
- •Programming becomes more interruptible; prompting replaces long deep-work blocks
- •Main challenge becomes ensuring quality, not producing lines of code
- •Prototypes are nearly free; experimentation becomes the default
- •Quality signals weaken: tests/docs can exist without “proof of use”
- 48:20 – 55:12
Simon’s AI stack: Claude Code (web/phone), GPT leapfrogging, and AI search
Simon outlines his day-to-day toolkit: Claude Code (especially hosted) for coding, GPT models as competitive alternatives, and AI-native search for research. He explains why ‘unsafe/YOLO’ modes change usability and why hosted environments can be safer for experimentation.
- •Claude Code for web enables phone-based coding and safer experimentation
- •‘Dangerously skip permissions’/YOLO mode makes agents dramatically more productive
- •GPT models increasingly match or exceed Claude depending on the month
- •AI search with citations has displaced direct Google usage for many queries
- 55:12 – 1:00:52
The pelican-riding-a-bicycle benchmark and the absurdity that reveals progress
Simon describes his playful SVG benchmark—pelican on a bicycle—as a qualitative way to compare models. Unexpectedly, better “pelicans” correlate with overall model capability, highlighting both real progress and the inherent ridiculousness of the space.
- •Benchmark targets LLM code generation via SVG, not image models
- •Qualitative comparisons can be more intuitive than numeric benchmark scores
- •Labs acknowledge and sometimes publicize pelican outputs
- •Humor and whimsy help keep perspective on what these systems are (and aren’t)
- 1:00:52 – 1:08:31
Hoarding know-how: building reusable personal libraries of tools and research
Simon explains how to accumulate leverage by saving working examples, experiments, and verified code—now easier than ever with agents. He shares his practice of maintaining public and private GitHub repos of tools and “coding agent research” to reuse as future context.
- •Career leverage comes from a backlog of tried techniques and working examples
- •GitHub repos as durable, searchable memory: tools + research projects
- •Value comes from code that was executed and verified, not untested “LLM reports”
- •Agents can search and combine prior artifacts to solve new problems quickly
- 1:08:31 – 1:16:30
Making agents reliable: red/green TDD, heavy tests, and project templates
Simon argues that tests are the key to trustworthy agentic engineering, and that agents should run code and verify behavior, not just generate it. He also recommends starting projects from templates so agents conform to preferred structure, style, and testing patterns from the first commit.
- •Core rule: agents must run tests; otherwise you’re back to copy/paste coding
- •Red/green TDD shorthand improves agent behavior and discipline
- •Verbose test suites become more acceptable when agents maintain them
- •Strong templates (even a single test) anchor style and structure for new repos
- 1:16:30 – 1:28:32
Prompt injection, the lethal trifecta, and why 97% security is a failing grade
The conversation turns to AI security: why prompt injection is an application-layer vulnerability and why common defenses are insufficient. Simon introduces the “lethal trifecta” (private data + malicious instructions + exfiltration) and predicts a “Challenger disaster” driven by normalization of deviance.
- •Prompt injection: untrusted text can override instructions in LLM-based apps
- •Lethal trifecta: private data + attacker-controlled input + ability to send data out
- •Defense mindset: assume attackers can make the agent do anything it’s allowed to do
- •Normalization of deviance risks a headline-scale failure; 97% detection isn’t enough
- 1:28:32 – 1:34:20
OpenClaw: huge demand for assistants, and an unresolved security nightmare
Simon explains OpenClaw’s meteoric rise and why it embodies the risky assistant people still desperately want. The central opportunity is building a version that preserves usefulness while preventing catastrophic data leaks and unsafe actions—something even major labs hesitated to ship.
- •OpenClaw’s rapid adoption shows massive demand for capable personal agents
- •It’s close to the ‘lethal trifecta’ scenario, creating serious exfiltration risk
- •Model competence mitigates some failures, but not reliably
- •Big opportunity: a ‘safe OpenClaw’ with constrained blast radius and stronger guarantees
- 1:34:20 – 1:38:11
What’s next for Simon: AI for journalism, a ‘not-a-book,’ and new work models
Simon shares his focus on open-source tools for data journalism and how AI can help journalists extract structured truth from messy sources—if treated as unreliable. He also discusses monetizing his blog and a preferred consulting style: “zero-deliverable consulting” that sells focused time, not artifacts.
- •Building AI-assisted tools to help journalists analyze documents and data
- •Journalists are well-suited to handle unreliable sources—AI included
- •Blog sponsorships and newsletters becoming meaningful income streams
- •Zero-deliverable consulting: paid expertise time without reports or code deliverables
- 1:38:11 – 1:39:50
Closing on an unexpected bright spot: Kakapo parrots are thriving
They end with rare good news: the Kakapo parrot in New Zealand is having a strong breeding season tied to rimu tree fruiting cycles. It’s a light, hopeful coda after a conversation about rapid technological change and risk.
- •Kakapo population is critically small but seeing new chicks
- •Breeding depends on rimu tree mass fruiting; 2026 is a good year
- •Webcams and public visibility make conservation progress tangible
- •A reminder to keep perspective amid AI acceleration and anxiety
