Lenny's PodcastClaude Code head Boris Cherny: Why he ships 30 PRs a day
Through hundred-percent AI-written code and parallel running agents on autopilot; 'clodify everything,' unlimited tokens, and latent demand make the builder.
CHAPTERS
- 0:00 – 3:55
Claude Code’s shock effect: 100% AI-written code and PRs at extreme velocity
Boris opens with a vivid picture of what “coding is solved” looks like in practice: he ships 10–30 PRs a day with code written entirely by Claude Code and runs multiple agents in parallel. He argues the job is becoming more enjoyable by removing minutiae, while pushing the industry toward a world where everyone can build software.
- •Boris hasn’t hand-edited a line of code since November; Claude writes everything
- •Multi-agent workflow: several agents running concurrently while he works
- •Engineer productivity and job satisfaction increase when the tedious parts disappear
- •Prediction: coding becomes broadly accessible; “software engineer” shifts toward “builder”
- •Early hint of the next shift: models moving from execution to proposing what to build
- 3:55 – 5:35
Why Boris left Anthropic for Cursor—and returned two weeks later
Boris recounts his brief move to Cursor, motivated by admiration for the product and team. He quickly realized he missed Anthropic’s mission-driven culture—especially the centrality of safety—which he needs personally to feel fulfilled at work.
- •Cursor’s team and product impressed him; they saw AI coding’s direction early
- •Anthropic’s mission (safety) was the decisive pull to return
- •Mission-driven alignment as a source of personal motivation and happiness
- •Anthropic’s culture: everyone can articulate why safety matters
- 5:35 – 7:12
One-year retrospective: Claude Code adoption, growth curves, and why it surprised them
The conversation zooms out to Claude Code’s rapid impact: a meaningful share of global GitHub commits and accelerating adoption. Boris emphasizes the most surprising part isn’t today’s scale, but the continued acceleration across metrics and the way user feedback keeps shaping the product.
- •Public stats (e.g., % of GitHub commits) undercount private repo usage
- •Growth continues accelerating rather than plateauing
- •Claude Code started as a “little hack,” not a guaranteed flagship product
- •User feedback loops are the dominant driver of improvements
- •Early external reception was slower—people needed time to understand the new paradigm
- 7:12 – 13:58
Origin story: the terminal, tool-use breakthrough, and the ‘under-resource to learn’ product lesson
Boris traces Claude Code’s beginnings: early prototypes, learning the model layer, and a pivotal tool-use moment where the model figured out how to answer a real-world question using bash. The team stayed terminal-first because it was the fastest form factor to iterate as model capabilities improved weekly.
- •Boris prototyped widely, then detoured into post-training to understand model behavior
- •Key “wow” moment: model learns to use bash tools without explicit instruction
- •Terminal-first wasn’t ideological—it was the fastest way to ship and iterate
- •Deliberate under-resourcing can force simplicity and speed early
- •Internal DAU chart went vertical after launch inside Anthropic
- 13:58 – 17:54
How fast AI is changing software: from 20–30% to 100% AI code, and what becomes the bottleneck
Boris explains how his own workflow crossed from partial AI assistance to total AI code generation over months, culminating in 100% AI-written code. He describes the new bottlenecks—review, correctness, safety—and how Anthropic uses Claude for automated code review with human checkpoints.
- •Progression: ~20–30% AI code early → 100% by November
- •Claude reviews 100% of PRs at Anthropic, with human review afterward
- •‘Hands-off’ isn’t fully safe yet for production code; verification still matters
- •Productivity gains are unprecedented compared to traditional DevEx improvements
- •The exponential mindset: tracing the trend line beats intuition
- 17:54 – 22:25
The next frontier: Claude as coworker—idea generation from feedback, telemetry, and bug reports
With code generation becoming reliable, Boris sees the frontier shifting to deciding what to build. He describes Claude increasingly proposing fixes and features by synthesizing internal feedback channels, telemetry, and bug reports—moving from assistant to proactive collaborator.
- •Claude starts to propose work items, not just implement them
- •Pointing Claude at Slack feedback threads produces actionable PRs
- •Product discovery and prioritization become more AI-assisted
- •Model quality improvements translate into better “what next?” suggestions
- •The ‘holy grail’ shifts from building → reviewing → choosing
- 22:25 – 24:21
Downsides of rapid innovation: staying current, mindset shifts, and ‘Claude can debug like you’
Boris highlights a subtle cost of fast-moving models: humans get anchored to old capabilities. He shares a story where he debugged a memory leak the traditional way while a newer teammate had Claude replicate the workflow, write analysis tooling, find the issue, and open a PR faster than he could.
- •Model capabilities evolve fast enough that habits quickly become outdated
- •Newer team members may be more ‘AGI-forward’ in approach than veterans
- •Claude can autonomously run classic debugging workflows (heap snapshots, analysis tools)
- •Humans must regularly recalibrate what’s now delegable to agents
- •Rapid change can create cognitive friction and workflow whiplash
- 24:21 – 27:58
Team principles: underfunding projects, moving today, and why ‘unlimited tokens’ unlocks innovation
Boris lays out operating principles for the Claude Code team: small teams ship faster, urgency beats perfection, and token budgets should be generous early. He argues token spend is often trivial relative to salaries, and the goal is to maximize learning before optimizing cost.
- •Underfunding can force ‘clodification’—automating work rather than adding headcount
- •A bias for speed: if it can be done today, do it today
- •Unlimited tokens as an innovation perk (freedom to explore ‘crazy’ ideas)
- •Optimize cost later (model choice, efficiency) after value is proven
- •Token spend can reach very high levels for some engineers as usage scales
- 27:58 – 36:02
Will coding skills still matter? A printing press analogy for democratized creation
Boris argues that understanding the ‘layer under the layer’ matters for now, but may fade in importance within a couple years. He compares AI coding to the printing press: a specialized skill (scribes) becomes democratized, unleashing new waves of creativity and societal change—while still causing disruption.
- •Coding knowledge remains helpful short-term, but may become less essential
- •Historical arc: switches → punch cards → software → higher abstractions → agents
- •Printing press analogy: volume up, cost down, literacy expands, new eras emerge
- •Engineers may shift from syntax to systems thinking, user needs, and coordination
- •Transition will be disruptive and painful even if long-term benefits are huge
- 36:02 – 40:42
Which roles transform next: agents for PM, design, and any computer-based work (Cowork)
Boris predicts AI’s impact spreads first to roles adjacent to engineering—PM, design, data science—then broadly to any computer-mediated job. He clarifies what “agent” should mean: an LLM that can use tools and act in systems (Docs, email, Slack), not just chat.
- •Next wave: roles adjacent to engineering, then most computer tool-based work
- •Definition of ‘agent’: tool-using, acting system—not merely conversational AI
- •Cowork as an on-ramp for non-engineers to experience agentic workflows
- •Anthropic’s broader effort: economists/policy/social impact to guide societal transition
- •The title/role boundaries blur; teams trend toward hybrid ‘builder’ profiles
- 40:42 – 47:29
Succeeding in the AI era: experimentation, becoming a generalist, and roles converging into ‘builder’
Boris’s advice centers on active experimentation and interdisciplinary breadth. He describes a team where everyone codes—including PM, design, finance, data science—and argues the most rewarded people will be AI-native generalists who can span user needs, business context, and execution.
- •Primary advice: don’t fear tools—use them constantly and stay on the frontier
- •Shift toward generalism and cross-functional fluency
- •On Claude Code team, everyone codes; specialization becomes a matter of degree
- •Role boundaries blur further; ‘software engineer’ may fade in favor of ‘builder’
- •Enjoyment trends differ by function; designers may experience more mixed outcomes
- 47:29 – 51:55
Latent demand: building where users are—and where the model ‘wants to go’
Boris explains latent demand as the key product principle: observe how users “misuse” a product to meet real needs, then build directly for that behavior. He extends the concept to models: instead of boxing an LLM into rigid workflows, expose it with minimal scaffolding and let it choose tools and sequence—staying ‘on distribution.’
- •Classic latent demand: discover real use by observing workaround behaviors
- •Examples: Facebook groups → Marketplace; profile views → Dating
- •Claude Code misuse for non-coding tasks signaled demand for Cowork
- •Modern latent demand: design around what the model is trying to do (on-distribution)
- •Inversion: ‘the product is the model’ + minimal scaffolding + tool access
- 51:55 – 1:08:39
Cowork built in 10 days—and Anthropic’s three-layer safety approach
Boris details how Cowork emerged from observed demand and shipped fast by reusing Claude Code inside the desktop app, largely built with Claude Code itself. He then outlines Anthropic’s safety stack: mechanistic interpretability/alignment, controlled evals, and real-world deployment learning—why ‘research preview’ releases matter.
- •Cowork shipped fast (10 days) by embedding Claude Code in desktop experience
- •Safety guardrails include sandboxing/VM approaches to constrain agent behavior
- •Release early to learn product fit and real-world safety behavior
- •Three safety layers: alignment/interpretability, evals, and in-the-wild observation
- •‘Race to the top’: open-sourcing safety components to raise industry standards
- 1:08:39 – 1:27:44
Pro tips and closing: plan mode, best model first, multi-interface workflows, Codex thoughts, and post-AGI miso
Boris shares practical usage tactics—use the most capable model, start in plan mode, and explore different interfaces (terminal, desktop, mobile, Slack). He briefly comments on Codex and competition, then ends with personal notes: Ukrainian roots with Lenny, and a post-AGI dream of making miso on long time scales.
- •Tip: best model can be cheaper overall due to fewer corrections and retries
- •Plan mode: force upfront reasoning; then auto-accept edits after agreement
- •Try multiple form factors; ‘Claude Code’ is the same agent across surfaces
- •Competition (Codex) is healthy; focus remains on user problems, not rivals
- •Post-AGI vision: return to rural life rhythms—making miso and thinking long-term
