Head of Claude Code: What happens after coding is solved | Boris Cherny
Boris Cherny explains how Claude Code has turned software engineering into a 200 % productivity boost, why latent demand drives product discovery, and what the next AI‑augmented roles will look like.
1h 27m·Guest Boris Cherny·Host Lenny Rachitsky·
AI Coding Revolution & Productivity Gains
Claude Code has moved the bulk of coding from manual typing to AI‑generated code, delivering massive productivity lifts and reshaping engineers' relationship with software. The shift is measurable in commit statistics, engineer output, and job satisfaction.
AI coding has eliminated manual code editing, boosting engineer productivity by 200 %
- Boris reports that he no longer edits a single line of code by hand and that productivity per engineer has risen two‑fold.
- The removal of minutiae allows engineers to focus on higher‑level design and problem‑solving.
- This change is reflected in daily shipping volumes of 10‑30 PRs per engineer.
- The productivity jump is corroborated by internal metrics showing a 200 % increase.
- Such gains free engineers to spend time on strategic work rather than rote implementation.
Quote
00:00:13“I have never enjoyed coding as much as I do today because I don't have to deal with all the minutia. Productivity per engineer has increased 200%.” — Boris Cherny
Quote
00:21:21“Productivity per engineer has increased 200%.” — Boris Cherny
AI‑generated code now accounts for a growing share of GitHub commits, projected to hit 20 % by year‑end
- Semi‑analysis reports that 4 % of all GitHub commits are already authored by Cloud Code.
- The trajectory suggests a fifth of all commits will be AI‑generated by the end of the year.
- This rapid adoption mirrors the speed at which developers have stopped writing code manually.
- The statistic underscores a broader industry shift toward AI‑assisted development.
- It also hints at future tooling and workflow changes for the entire ecosystem.
Quote
00:06:01“4% of all GitHub commits are authored by cloud code now. and they predicted it'll be a fifth of all code commits on GitHub by the end of the year.” — Lenny Rachitsky
Quote
00:06:10“The day that we're recording this, Spotify just put out this headline that their best developers haven't written a line of code since December thanks to AI.” — Lenny Rachitsky
Engineers report higher job satisfaction with AI assistance, with 70 % enjoying their work more
- Lenny’s poll shows that 70 % of engineers say AI tools make their jobs more enjoyable.
- Only about 10 % report a decrease in satisfaction, indicating a net positive impact.
- Designers show a slightly lower uplift, suggesting role‑specific variance.
- The data aligns with anecdotal reports from Anthropic staff who love the new workflow.
- Increased enjoyment is linked to reduced drudgery and more creative focus.
Quote
00:44:38“70% of engineers are enjoying their job more since adopting AI tools.” — Lenny Rachitsky
Quote
00:45:01“About 10% said they're enjoying their job less.” — Lenny Rachitsky
Latent Demand and Product Discovery
Anthropic builds products by watching how users repurpose existing tools, turning unexpected use‑cases into new features. This “latent demand” mindset lets the team surface high‑impact opportunities before they are even imagined.
Latent demand drives product direction: building tools where users already adapt them
- Boris describes latent demand as bringing a tool to where people are and making their existing workflows easier.
- The terminal‑first design of Quad Code was surprising but matched where power users already worked.
- By observing users hacking the product, the team identifies natural extensions.
- This approach reduces friction and accelerates adoption.
- It also creates a virtuous loop: users discover new uses, the team builds them, more users adopt.
Quote
00:12:15“Latent demand where we bring the tool to where people are and it makes existing workflows a little bit easier, but also because it's in a terminal. It's a little surprising.” — Boris Cherny
Quote
00:47:29“Latent demand is the single most important principle in product.” — Boris Cherny
Facebook Marketplace emerged from users repurposing groups for buying/selling, illustrating latent demand
- Early Facebook groups were unintentionally used for commerce, with 40 % of posts being buy/sell.
- Recognizing this, the team built dedicated Marketplace groups, then a full Marketplace product.
- The rapid uptake proved the power of surfacing latent demand.
- This case study shows that observing user hacks can reveal whole new product categories.
- Anthropic applies the same lens to AI tools, turning hacks into features.
Quote
00:48:06“Facebook Marketplace started based on the observation that 40% of posts in Facebook groups are buying and selling stuff.” — Boris Cherny
Quote
00:48:30“It was obvious if we build a better product to let people buy and sell, they're going to like it.” — Boris Cherny
Co‑work originated from non‑technical users employing Quad Code for tasks like tomato planting, prompting a dedicated product
- Users began using Quad Code to control a tomato‑growing robot, analyze genomes, and recover photos.
- These non‑coding hacks highlighted a demand for a more user‑friendly interface.
- Within weeks the team built a desktop app (Co‑work) to surface the same capabilities.
- The product now supports dozens of non‑technical use‑cases, confirming the latent‑demand hypothesis.
- This evolution illustrates how unexpected user behavior can seed entirely new product lines.
Quote
00:49:20“There were people using Quad Code to grow tomato plants, analyze their genome, recover photos from a corrupted hard drive.” — Boris Cherny
Quote
00:52:14“Co‑work was built in 10 days using Quad Code.” — Boris Cherny
Model Safety and Alignment
Safety is Anthropic’s core mission. The team layers alignment work from low‑level neuron monitoring to real‑world deployment safeguards, releasing products early to test safety in the wild.
Safety is Anthropic's core mission; internal culture constantly emphasizes it
- Boris left Cursor because Anthropic’s safety‑first mission resonated with his values.
- Within Anthropic, every hallway conversation circles back to safety.
- The mission drives hiring, research priorities, and product decisions.
- Safety is framed as a non‑negotiable pillar, not an after‑thought.
- This cultural focus enables rapid yet responsible product iteration.
Quote
00:05:35“The thing that drew me to Anthropic was the mission. And it was, you know, it's all about safety.” — Boris Cherny
Quote
00:05:44“When you talk to people at Anthropic, if you ask them why they're here, the answer is always safety.” — Boris Cherny
Safety work spans alignment, mechanistic interpretability, lab testing, and real‑world monitoring
- The team uses three safety layers: alignment/mechanistic interpretability, controlled lab experiments, and wild‑deployment observation.
- They can monitor specific neurons for deceptive behavior and intervene.
- Lab‑scale petri‑dish experiments let them probe model responses under synthetic conditions.
- Real‑world monitoring catches failures that lab tests miss, ensuring continuous safety improvement.
- This multi‑layered approach balances rigor with practical deployment speed.
Quote
00:22:33“The lowest level is alignment and mechanistic interpretability. We can monitor neurons related to deception.” — Boris Cherny
Quote
00:55:31“The third layer is seeing how the model behaves in the wild.” — Boris Cherny
Early release of Cloud Code as a research preview allowed rigorous safety testing before public launch
- Cloud Code was used internally for 4‑5 months before public release to evaluate safety.
- The early release acted as a research preview, giving the team real user data while retaining control.
- Open‑source sandbox environments let external parties test agents safely.
- Feedback from internal and external usage informed safety mitigations before scaling.
- This strategy demonstrates how early, controlled exposure can accelerate safe AI deployment.
Quote
00:55:45“We released cloud code early as a research preview to study safety.” — Boris Cherny
Quote
00:58:45“We open‑source a sandbox so any agent can run safely.” — Boris Cherny
Tool‑Using Agents and Beyond Coding
Claude Code agents now act on the world: they can send emails, browse the web, and automate non‑coding tasks. This expands AI’s role from code generation to general‑purpose digital assistants.
Quad Code agents now act on the world: they can send emails, use Slack, and operate browsers
- Boris explains that agents can invoke Gmail, Slack, and other tools autonomously.
- This marks a shift from pure code generation to acting as a co‑worker.
- The ability to chain tool use enables complex workflows without human intervention.
- Early examples include fetching music metadata and filing tickets.
- As models improve, the range of actionable tools will expand dramatically.
Quote
00:08:34“It can actually use your Gmail, it can use your Slack, it can do all these things for you and it's quite good at it.” — Boris Cherny
Quote
00:08:18“The AI doesn't just write the code. It's not just a conversation partner, but it actually uses tools. It acts in the world.” — Boris Cherny
Co‑work extends this capability to non‑technical workflows like paying parking tickets and filling forms
- Co‑work can automate mundane tasks such as paying parking tickets, completing PDFs, and summarizing emails.
- Users report that the agent can fill medical forms and handle browser‑based interactions.
- This demonstrates that AI agents are useful beyond software engineering.
- By abstracting tool use, non‑technical users can leverage AI for everyday productivity.
- The rapid improvement of the model makes these capabilities increasingly reliable.
Quote
00:18:40“I just had Co‑work do it. It can actually pay a parking ticket the other day.” — Boris Cherny
Quote
00:35:13“Co‑work can fill out medical forms automatically.” — Boris Cherny
Future AI agents will handle any computer‑tool‑based job, making roles like product manager or designer increasingly AI‑augmented
- Boris predicts that adjacent roles (product, design, data science) will be the next wave of AI augmentation.
- Agents that can use spreadsheets, Slack, and email will automate many routine aspects of these jobs.
- The core skill will shift from manual execution to prompting and supervising agents.
- This will blur traditional role boundaries, creating hybrid “builder” positions.
- The trajectory suggests that most computer‑based work will eventually be AI‑driven.
Quote
00:36:24“I think it's going to be a lot of the roles that are adjacent to engineering.” — Boris Cherny
Quote
00:36:45“Any job that uses computer tools in this way will be next.” — Boris Cherny
Token Economics and Model Selection
Choosing the right model and token policy dramatically impacts cost and speed. Anthropic encourages generous token budgets to accelerate experimentation, while emphasizing the value of using the most capable model.
Using the most capable model (Opus 4.6) reduces token consumption and cost despite higher per‑token price
- Opus 4.6 completes tasks with fewer tokens than older, cheaper models.
- Fewer tokens mean less hand‑holding and fewer correction cycles.
- The net cost is lower even though the per‑token price is higher.
- Boris advises always enabling “maximum effort” on the top model.
- This approach yields faster, cleaner outputs and overall savings.
Quote
01:09:27“Use the most capable model. Currently that's Opus 4.6. I have maximum effort enabled always.” — Boris Cherny
Quote
01:09:35“It's actually cheaper if you use the most capable model because it does the same thing with less correction.” — Boris Cherny
Giving engineers abundant tokens encourages experimentation and accelerates innovation
- Boris recommends giving engineers as many tokens as possible, even unlimited tokens at some companies.
- A generous token budget lets teams try bold ideas without worrying about cost.
- Early experiments that would have been “too crazy” become feasible.
- The policy has led to rapid discovery of new use‑cases.
- Token generosity is framed as a strategic investment rather than a cost center.
Quote
01:26:13“Give engineers as many tokens as possible. Some companies even offer unlimited tokens.” — Boris Cherny
Quote
01:26:55“The advice is to be loose with your tokens; the cost is secondary to innovation.” — Boris Cherny
Optimizing for token cost too early can hinder progress; better to wait for next model improvements
- Boris warns against early cost‑cutting, suggesting that waiting for a better model yields higher ROI.
- Scaffolding can improve performance 10‑20 % but is quickly eclipsed by a new model.
- Over‑optimizing token usage can stifle experimentation.
- The “bitter lesson” reinforces betting on more general, future models.
- Teams should prioritize speed of learning over marginal token savings.
Quote
01:05:30“Don't try to optimize cost early. Give engineers tokens to experiment.” — Boris Cherny
Quote
01:05:38“Scaffolding gives 10‑20 % gains but gets wiped out by the next model.” — Boris Cherny
Early Release, Feedback Loops, and Iteration
Anthropic’s product process emphasizes ultra‑fast feedback: user reports become PRs within minutes, and multiple agents run in parallel to offload work, enabling rapid iteration.
Rapid feedback cycles (minutes) turn user reports into immediate product improvements
- When a user submits feedback, the team can ship a fix within 5 minutes.
- This speed makes users feel heard and encourages more feedback.
- The loop creates a virtuous cycle of improvement and adoption.
- Boris cites the DAU chart spiking after early rapid fixes.
- Fast iteration is a core competitive advantage for Claude Code.
Quote
00:19:51“I would just go in and fix every single thing as fast as possible, within a minute or five.” — Boris Cherny
Quote
00:11:46“The chart just went vertical pretty immediately after we fixed things fast.” — Boris Cherny
Multi‑quading (running many agents in parallel) enables developers to offload tasks and focus on higher‑level work
- Boris describes “multi‑quading” as running dozens of agents simultaneously.
- Engineers can have agents handle PR reviews, project‑management, and data‑analysis.
- This frees humans to think strategically rather than manually executing repetitive steps.
- The ability to run agents for hours or weeks further reduces human supervision.
- Multi‑quading turns AI into a true co‑worker, not just a code generator.
Quote
00:47:21“We call this multi‑quading – running many Quad sessions in parallel.” — Boris Cherny
Quote
01:12:45“Co‑work can run a bunch of quads in parallel, so I can get coffee while it works.” — Boris Cherny
Launching Co‑work in ten days demonstrated the power of building on top of existing agents
- The team built Co‑work in just ten days by reusing Quad Code as the core engine.
- This rapid development proved the modularity of the agent platform.
- Early launch allowed real‑world feedback that shaped the product quickly.
- The speed of delivery showcases Anthropic’s “release early, iterate fast” philosophy.
- Co‑work’s success validates the latent‑demand approach and the underlying agent architecture.
Quote
00:52:14“Co‑work was built in 10 days using Quad Code.” — Boris Cherny
Quote
00:53:03“It was built in ten days, and we launched it early to get feedback.” — Boris Cherny
The Bitter Lesson: General vs Specialized Models
Across AI research, broader, more general models consistently outperform narrow, task‑specific ones. Anthropic embraces this by betting on future, larger models rather than fine‑tuning current versions.
General models consistently outperform narrow, task‑specific models across domains
- Boris cites Rich Sutton’s “bitter lesson”: the most general solution wins.
- Evidence comes from self‑driving cars, language tasks, and coding agents.
- Specialized models give diminishing returns as general models improve.
- The principle guides Anthropic’s roadmap to prioritize model scaling.
- It also informs product design: build abstractions that work with any model.
Quote
01:04:45“The more general model will always outperform the more specific model.” — Boris Cherny
Quote
01:05:12“Bet on the more general model and don't fine‑tune narrow ones.” — Boris Cherny
Betting on future, more general models yields higher long‑term ROI than fine‑tuning current models
- Boris advises building for the model six months out, not the current one.
- Early prototypes may be weak, but once the model improves, product‑market fit accelerates.
- This strategy avoids wasted effort on features that become obsolete.
- It aligns engineering timelines with the AI development curve.
- Companies that wait for the next model often capture market share faster.
Quote
01:06:25“Build for the model six months out; that's when the product will click.” — Boris Cherny
Quote
01:05:38“Scaffolding gives 10‑20 % gains but gets wiped out by the next model.” — Boris Cherny
Scaffolding around models gives diminishing returns as model capabilities jump
- Adding elaborate pipelines can improve performance 10‑20 %.
- However, a newer model release often eclipses those gains entirely.
- Therefore, investing heavily in scaffolding can be wasteful.
- The team prefers to wait for the next model and keep the system lightweight.
- This philosophy reduces technical debt and speeds up iteration.
Quote
01:05:38“Scaffolding can improve performance 10‑20 % but gets wiped out by the next model.” — Boris Cherny
Quote
01:05:44“We should wait for the next model rather than over‑optimizing now.” — Boris Cherny
Career Advice: Generalist Skills and Token Generosity
The most valuable engineers blend product, design, and infrastructure expertise, thrive in psychologically safe environments, and apply common‑sense thinking over rigid processes.
Engineers who blend product, design, and infrastructure skills become most valuable
- Boris notes that on his team, product managers, designers, data scientists all code.
- Hybrid roles (product‑infrastructure, product‑design) enable better problem framing.
- Such engineers can translate user needs directly into technical solutions.
- The cross‑disciplinary fluency accelerates feature delivery and reduces hand‑offs.
- Companies that cultivate these hybrids gain a competitive edge in AI‑augmented product cycles.
Quote
01:41:33“Our product manager codes, our engineering manager codes, our designer codes, our finance guy codes.” — Boris Cherny
Quote
01:42:40“The strongest engineers are hybrid product and infrastructure engineers.” — Boris Cherny
Psychological safety and tolerance for failure fuel innovation
- Boris emphasizes that 80 % of ideas are bad; teams must cut losses quickly.
- A safe environment lets people experiment without fear of blame.
- Failure is treated as data, not a career risk.
- This culture leads to rapid iteration and breakthrough ideas.
- Psychological safety is as important as technical skill for AI product teams.
Quote
01:15:21“There's no roadmap for innovation. You have to give people psychological safety that it's okay to fail.” — Boris Cherny
Quote
01:15:38“If the idea is bad you cut your losses and move on.” — Boris Cherny
Common sense and first‑principles thinking outweigh rigid processes in product building
- Boris advises teams to develop their own common sense rather than blindly follow processes.
- He stresses asking “why” and building from first principles.
- Over‑reliance on templates can stifle creativity.
- Common sense helps spot bad ideas early and avoid wasted effort.
- This mindset is especially critical when building AI‑augmented products where the landscape shifts fast.
Quote
01:24:14“Use common sense. Think from first principles.” — Boris Cherny
Quote
01:24:20“People fail when they follow a process without thinking about it.” — Boris Cherny
Historical Analogy: Printing Press and Democratization
The printing press multiplied information output 100× and lowered costs, unlocking mass literacy. Anthropic sees AI as a similar democratizing force for code creation, promising a future where programming is universal.
The printing press multiplied information output 100× and reduced cost, analogous to AI democratizing code creation
- Boris cites that printed material grew 100× in 50 years after Gutenberg.
- Costs fell by roughly two orders of magnitude, making books affordable.
- This massive scale‑up mirrors the rapid adoption of AI‑generated code.
- Just as the press unlocked the Renaissance, AI could unlock a new era of software creation.
- The analogy underscores the societal impact of making a formerly elite skill universal.
Quote
01:33:27“The volume of printed material just went way up. It went down something like 100x over the next 50 years.” — Boris Cherny
Quote
01:33:44“The printing press multiplied printed material 100× and reduced cost dramatically.” — Boris Cherny
Literacy rose from <1 % to 70 % over centuries; AI will similarly lift programming literacy globally
- In the 1400s, less than 1 % of the population could read or write.
- Over 200 years, literacy climbed to about 70 % worldwide.
- The same exponential diffusion is expected for AI‑assisted programming.
- As tools become easier, more people will be able to build software.
- This shift will reshape education, work, and economic opportunity.
Quote
01:33:04“In the mid‑4000s literacy was sub‑1 % of the population.” — Boris Cherny
Quote
01:34:01“Over the next 200 years it went up to about 70 % globally.” — Boris Cherny
Early adopters of new tools (scribes) felt excitement similar to modern engineers embracing AI agents
- Scribes in the 15th century were a tiny elite who loved freeing themselves from copying.
- They relished the new ability to focus on illustration and binding.
- Today, engineers feel similar joy when AI removes tedious coding.
- The parallel highlights how technology can shift the nature of creative work.
- Recognizing this pattern helps anticipate cultural resistance and adoption curves.
Quote
01:34:20“The thing that I do like is drawing the art in books and then doing the book binding. I'm glad my time is freed up.” — Boris Cherny
Quote
00:00:13“I love coding more because I don't have to deal with all the minutia.” — Boris Cherny
Non‑Technical Use Cases and Co‑Work Applications
Co‑work extends AI assistance to everyday tasks, from paying tickets to filling medical forms, and is available across platforms, making AI a ubiquitous personal assistant.
Co‑work can automate mundane admin tasks: paying tickets, filling PDFs, summarizing emails
- Users have used Co‑work to pay parking tickets automatically.
- The agent can fill out medical forms and other repetitive PDFs.
- Email summarization and reply generation are built‑in features.
- These capabilities free non‑technical users from routine bureaucracy.
- The speed and accuracy have improved dramatically as the model matured.
Quote
00:18:40“I just had Co‑work do it. It can actually pay a parking ticket the other day.” — Boris Cherny
Quote
00:35:13“Co‑work can fill out medical forms automatically.” — Boris Cherny
Running many agents in parallel lets users treat AI as a virtual assistant across projects
- Multi‑quading enables dozens of agents to run simultaneously.
- Engineers can have one agent manage project status, another handle email, another run data analysis.
- This parallelism turns AI into a team of assistants rather than a single tool.
- Users report being able to step away (e.g., get coffee) while agents work unattended.
- The approach scales to weeks‑long tasks without human supervision.
Quote
00:47:21“We call this multi‑quading – running many Quad sessions in parallel.” — Boris Cherny
Quote
01:12:45“I can have as many tasks running as I want and then go get a coffee while it runs.” — Boris Cherny
Integrations across iOS, Android, desktop, Slack, and web make AI assistance ubiquitous
- Claude Code is available on mobile apps, a desktop client, a web UI, and IDE extensions.
- Slack integration lets agents act directly in chat channels.
- The breadth of platforms ensures users can stay in their preferred workflow.
- This ubiquity lowers the barrier to adoption for non‑technical users.
- Anthropic continues to add form‑factors to keep the model close to the user.
Quote
00:12:57“Quad Code is available in the iOS and Android app, desktop app, website, and IDE extensions in Slack and GitHub.” — Boris Cherny
Quote
01:22:08“Co‑work has a Chrome extension that lets you talk to Claude while browsing.” — Boris Cherny
⚙ Agent-readable JSON index — click to expand
{ "memcast_version": "0.1", "episode": { "id": "We7BZVKbCVw", "title": "Head of Claude Code: What happens after coding is solved | Boris Cherny", "podcast": "Lenny's Podcast", "guest": "Boris Cherny", "host": "Lenny Rachitsky", "source_url": "https://www.youtube.com/watch?v=We7BZVKbCVw", "duration_minutes": 88 }, "concepts": [ { "id": "ai-coding-revolution-productivity-gains", "title": "AI Coding Revolution & Productivity Gains", "tags": [] }, { "id": "latent-demand-and-product-discovery", "title": "Latent Demand and Product Discovery", "tags": [] }, { "id": "model-safety-and-alignment", "title": "Model Safety and Alignment", "tags": [ "ai-safety" ] }, { "id": "tool-using-agents-and-beyond-coding", "title": "Tool‑Using Agents and Beyond Coding", "tags": [ "automation", "tool-use" ] }, { "id": "token-economics-and-model-selection", "title": "Token Economics and Model Selection", "tags": [] }, { "id": "early-release-feedback-loops-and-iteration", "title": "Early Release, Feedback Loops, and Iteration", "tags": [] }, { "id": "the-bitter-lesson-general-vs-specialized-models", "title": "The Bitter Lesson: General vs Specialized Models", "tags": [ "bitter-lesson" ] }, { "id": "career-advice-generalist-skills-and-token-generosity", "title": "Career Advice: Generalist Skills and Token Generosity", "tags": [ "team-culture" ] }, { "id": "historical-analogy-printing-press-and-democratization", "title": "Historical Analogy: Printing Press and Democratization", "tags": [] }, { "id": "non-technical-use-cases-and-co-work-applications", "title": "Non‑Technical Use Cases and Co‑Work Applications", "tags": [] } ] }