Lenny's Podcast: Product | Career | Growth

Head of Claude Code: What happens after coding is solved | Boris Cherny

February 19, 2026

Key Takeaways Copied to clipboard!

  • Boris Cherny believes coding is virtually solved, evidenced by his personal workflow where 100% of his code is written by Claude Code (Quad Code) since November, leading to a 200% increase in productivity per engineer at Anthropic. 
  • The success of Claude Code and Cowork is driven by the product principle of latent demand, observed both in engineers using the terminal tool for non-coding tasks and in the model's inherent capabilities, leading to the creation of Cowork in just 10 days. 
  • Boris advises that to foster innovation, teams should initially underfund projects slightly and give engineers unlimited tokens to encourage rapid experimentation, only optimizing for cost once an idea proves successful. 
  • Cowork achieved immediate success, unlike Claude Code's initial slow adoption, by capitalizing on latent demand observed from non-technical uses of Claude Code and being built rapidly in just 10 days using Claude Code itself. 
  • Anthropic employs a three-layered approach to AI safety—mechanistic interpretability, controlled evals, and real-world observation—to study and ensure model safety, often releasing products like Claude Code early to gather crucial third-layer feedback. 
  • Successful AI product builders should bet on the general model six months out, avoid over-curating the model with strict workflows, and instead provide tools and goals, adhering to 'The Bitter Lesson' that general models outperform specific ones over time. 

Segments

Boris’s Coding Workflow
Copied to clipboard!
(00:00:00)
  • Key Takeaway: Boris Cherny’s entire code output is generated by Claude Code, with no manual editing since November, resulting in 10-30 pull requests daily.
  • Summary: Boris has not edited a line of code by hand since November, relying entirely on Claude Code (Quad Code) for his prolific output of 10 to 30 pull requests daily. He finds this method more enjoyable as it eliminates tedious minutiae, increasing his personal productivity significantly. He still reviews the AI-generated code and utilizes Claude Code for automatic code review on all pull requests.
Claude Code’s Rapid Growth
Copied to clipboard!
(00:05:35)
  • Key Takeaway: Claude Code grew from a terminal-based hack to authoring 4% of all public GitHub commits, with daily active users doubling in the last month alone.
  • Summary: The growth of Claude Code has been unexpectedly massive, reaching 4% of public GitHub commits, with predictions suggesting it could hit one-fifth of all commits by year-end. The initial prototype was intended as a simple hack, evolving from the Anthropic Labs team’s trajectory of building models good at coding, then tool use, then computer use. The product’s success was not immediate upon external release in February, taking months for broader understanding and adoption.
Innovation Through Experimentation
Copied to clipboard!
(00:15:01)
  • Key Takeaway: Major AI innovations, like Claude Code, often stem from giving teams psychological safety to fail and encouraging rapid iteration rather than forcing a strict roadmap.
  • Summary: Innovation requires space and psychological safety where failure is acceptable, allowing individuals to pull on promising threads like Boris did with the terminal-based Claude Code prototype. Although the initial internal reaction was muted, the rapid feedback cycle from users encouraged continuous improvement, leading to the product’s eventual success. Boris emphasizes that the initial advantage in a crowded market was speed, achieved by doing things today rather than waiting.
The Next Frontier Beyond Coding
Copied to clipboard!
(00:17:32)
  • Key Takeaway: The next frontier for AI agents is moving beyond coding, where Claude Code is already assisting in proactive idea generation based on telemetry and bug reports, and Cowork handles general tasks like project management.
  • Summary: Since coding is considered largely solved for Boris’s use cases, the focus is shifting to adjacent tasks where agents can proactively suggest bug fixes or new features by analyzing feedback and telemetry. Boris uses Cowork daily for non-coding tasks, including paying parking tickets and managing team project workflows. This expansion shows the transition from AI as a coding partner to an agent that acts in the world using tools like Slack and email.
Principles for AI Team Success
Copied to clipboard!
(00:24:02)
  • Key Takeaway: Key principles for the Claude Code team include underfunding initial efforts to force reliance on AI automation and encouraging engineers to use unlimited tokens to explore radical ideas before optimizing costs.
  • Summary: Underfunding teams slightly forces engineers to leverage AI tools (Quadify) to achieve high output, which is more effective than premature cost-cutting. Boris strongly advises giving engineers unlimited tokens initially, as the cost of experimentation at a small scale is low relative to salary, allowing for the discovery of breakthrough ideas. Optimization should only occur after a successful idea scales and token costs become significant.
Coding Skills and Historical Analogs
Copied to clipboard!
(00:27:58)
  • Key Takeaway: Boris predicts that the need for traditional coding skills will diminish within a year or two, comparing the current shift to the democratizing impact of the printing press on literacy.
  • Summary: Boris does not miss writing code because he views programming as a practical tool for building things, not an end in itself, and he is enjoying focusing on higher-level problem-solving. He compares the current transformation to the printing press, which drastically increased the volume of written material and eventually led to widespread literacy. While the transition will be disruptive, the ultimate outcome is a world where everyone can program, unlocking unimaginable new capabilities.
Advice for Succeeding in the AI Era
Copied to clipboard!
(00:40:42)
  • Key Takeaway: To succeed in the AI era, individuals should experiment deeply with AI tools and strive to become cross-disciplinary generalists rather than narrowly specialized engineers.
  • Summary: The most effective people will be AI-native generalists who are curious and cross over multiple disciplines, such as engineers with strong design sense or product managers who can code. Boris notes that on his team, everyone codes, blurring traditional role lines, and predicts the title ‘software engineer’ may soon be replaced by ‘builder.’ This shift allows engineers to focus on high-value activities like user collaboration and system thinking, which were previously obscured by tedious details.
Latent Demand and Product Building
Copied to clipboard!
(00:46:32)
  • Key Takeaway: Latent demand is identified by observing how users ‘misuse’ a product to solve an unaddressed need, which strongly signals where to build the next successful, purpose-built product.
  • Summary: Latent demand is revealed when users jump through hoops, like using the terminal-based Claude Code for non-technical tasks such as analyzing genomes or growing tomatoes, indicating a strong underlying need. This observation led directly to the creation of Cowork, which was built in 10 days to serve this non-technical user base immediately. A modern framing of latent demand also involves observing what the AI model itself is trying to do, rather than just what the user is doing.
Cowork’s Rapid Success and Latent Demand
Copied to clipboard!
(00:51:54)
  • Key Takeaway: Cowork was built in 10 days using Claude Code after exploring non-technical uses of the coding agent revealed latent demand for a desktop application.
  • Summary: Cowork launched immediately as a hit, contrasting with Claude Code’s slower initial adoption, because it addressed an already present, non-technical user need. The team built the entire product, including sophisticated security guardrails and a virtual machine, in about 10 days using Claude Code. Releasing products like Cowork early, even when rough, is essential for learning about latent demand and safety in the wild.
Anthropic’s Three Layers of Safety
Copied to clipboard!
(00:54:29)
  • Key Takeaway: AI safety is studied across three layers: mechanistic interpretability (observing internal model thinking), controlled evals (laboratory testing), and real-world behavior observation.
  • Summary: The lowest layer of safety involves alignment and mechanistic interpretability, allowing researchers to monitor specific neurons related to concepts like deception within the model. The second layer consists of controlled evals where the model is tested in synthetic, petri-dish-like situations for alignment. The third, and increasingly critical layer, is observing how the model behaves once released into the wild, which informs subsequent safety improvements.
Agent Anxiety and Multi-Interface Workflow
Copied to clipboard!
(00:59:38)
  • Key Takeaway: Boris mitigates agent anxiety by running multiple agents constantly across desktop, web, and mobile interfaces, making him less locked into any single terminal session.
  • Summary: Boris experiences low anxiety regarding blocked agents because he typically has five or more agents running concurrently across different platforms. He notes that his coding workflow is now split roughly equally between the terminal, the desktop app, and the surprisingly useful iOS app. This shift reflects a change where ‘coding’ is becoming more about describing intent than writing syntax, echoing historical shifts in programming methods.
Advice for Building AI Products
Copied to clipboard!
(01:03:22)
  • Key Takeaway: Effective AI product building requires avoiding strict orchestration, giving models tools to find context, betting on general models over fine-tuning, and building for the model six months in the future.
  • Summary: Do not box the model in by layering strict, step-by-step workflows; instead, provide tools and a goal, allowing the general model to figure out the execution path. This aligns with ‘The Bitter Lesson,’ suggesting one should always bet on the more general model, as scaffolding gains are often wiped out by the next model release. Building for the model six months out means accepting initial discomfort but ensuring the product clicks when future capabilities, like extended runtimes, arrive.
Pro Tips for Using Claude Code
Copied to clipboard!
(01:08:42)
  • Key Takeaway: Users should consistently use the most capable model (Opus 4.6), start tasks in plan mode to align on strategy before execution, and experiment across all available interfaces (terminal, desktop, mobile).
  • Summary: Using the most capable model is often cheaper overall because it requires less hand-holding and fewer tokens to complete complex tasks correctly. Plan mode, which simply instructs the model not to write code yet, is crucial for aligning on the strategy before execution, after which auto-accepting edits is recommended for one-shot completion. Engineers should explore all form factors, including the desktop and mobile apps, as the underlying Claude Code agent remains the same across all interfaces.
Post-AGI Plans and Life Philosophy
Copied to clipboard!
(01:11:20)
  • Key Takeaway: Boris’s post-AGI plan, or his preferred non-engineering activity, involves making miso, which teaches patience and thinking on multi-year time scales.
  • Summary: Boris previously lived in rural Japan, where he learned to make miso, a process requiring three months to several years for fermentation. This activity contrasts sharply with engineering’s fast pace by forcing him to think in long time scales. He stated that if he were not at Anthropic, he would likely be focusing on perfecting his miso-making.
Lightning Round Insights
Copied to clipboard!
(01:16:17)
  • Key Takeaway: Recommended technical reading includes ‘Functional Programming in Scala’ for its elegance, while ‘Accelerando’ by Charles Stross captures the current moment’s accelerating pace.
  • Summary: Boris highly recommends ‘Functional Programming in Scala’ for its foundational concepts, even if Scala itself is not used, as it shapes how he thinks about coding. He praises ‘Accelerando’ for capturing the feeling of rapid technological acceleration leading toward a singularity. He also enjoys the short stories of Cixin Liu, such as ‘The Wandering Earth,’ for offering a different perspective than Western science fiction.