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

Lenny's Podcast · video · 2026-02-19 source ↗ Reader 🖥️

practical · mixed — A long-form conversational interview (Lenny Rachitsky interviewing Boris Cherny, head of Claude Code at Anthropic): autobiographical narrative, practical advice for builders/engineers, and predictive claims about the future of software engineering roles. The Reader document already contained the full real dialogue (confirmed, not show notes); this is YouTube auto-generated captions, which contain recognizable ASR errors (e.g. 'Claude' rendered as 'Quad', 'Cloud', or once as 'GPT-4'; 'Claude Code' as 'Quad Code' or 'Cloud Code') that have been normalized to 'Claude'/'Claude Code' in this extraction's quotes.

Structure

Key terms

• latent demand
  (Boris) The product-design principle of watching what users are already improvising a product into doing despite it not being designed for that, and building directly for it — extended by Boris beyond human users to models themselves: watching what a model tries to do on its own ('being on distribution') and designing minimal scaffolding around that.
• underfunding
  (Boris) Deliberately assigning fewer people to a project than the work seems to need, as a management tactic — the resulting urgency and intrinsic drive to ship pushes engineers to delegate maximally to Claude rather than to ask for more headcount.
• build for the model six months out
  (Boris) A deliberate product-strategy bet: design for a more capable future model rather than optimizing for the current model's limitations, accepting weak product-market fit in the interim, because scaffolding built around today's limitations is typically wiped out by the next model generation anyway.
• plan mode Claude Code
  (Boris) Deliberately described as mechanically trivial: a single injected sentence telling the model not to write code yet, used for roughly 80% of Boris's own tasks as a lightweight checkpoint where the model proposes a plan before the human approves execution.
• race to the top
  (Boris, internal Anthropic framing) Open-sourcing safety-relevant infrastructure (e.g. Claude Code's agent sandbox, made to work with any agent) so that competing labs and vendors adopt the same safety practice, framed as a deliberate competitive dynamic Anthropic wants to win by raising the industry's safety floor rather than by keeping a proprietary safety advantage.

Propositions

• explicitDespite Claude authoring 100% of his code since November, a human review layer still exists over AI-generated work at Anthropic: Claude performs automatic review on 100% of pull requests, but a human layer still follows that, except for pure, throwaway prototype code that will never run anywhere.conf: high
  extendsWhen AI Writes Almost All Code, What Happens To Software Engineering? · p3 — This interview is the primary source behind that article's citation of Boris's '100% AI-written code' claim, and adds the crucial qualifying detail the secondary article doesn't carry: Claude also auto-reviews 100% of PRs, but a human layer still follows except for throwaway prototype code.
• explicitOver roughly the year since Claude Code's launch, per-engineer productivity on Boris's team increased about 200% while the engineering team grew roughly fourfold — a far larger gain than the 'a few percentage points per year' productivity improvements Boris previously saw leading code-quality efforts across Facebook, Instagram, and WhatsApp.conf: high
• explicitDeliberately 'underfunding' a project — assigning fewer people than the work seems to need — forces engineers to delegate maximally to Claude and ship faster, because the resulting urgency and intrinsic motivation to ship drives heavier AI use more effectively than adding headcount would.conf: high
  supportsDesigning the AI-native engineering organization · p3 — Boris's deliberate 'underfunding' practice is a concrete mechanism for exactly the claim that effective AI-era teams are smaller because the bottleneck shifts from building to alignment/decision-making — intentional scarcity is what forces the small team to lean maximally on Claude.
• explicitcontestedTeams should not try to optimize or cut AI token costs early; the priority should be giving engineers effectively unlimited tokens so they can try ideas that would otherwise seem too expensive to attempt, and only cost-optimize (e.g. by downgrading to a cheaper model) once a proven idea needs to scale.conf: high
  contradictsDesigning the AI-native engineering organization · p6 — Boris explicitly advises against managing AI token costs with early rigor ('don't try to optimize and cost-cut at the beginning'), while the other extraction advises managing AI token costs with the same rigor as cloud infrastructure costs from the start — a genuine, useful tension about when cost discipline should begin, not just whether it should exist.
• explicitThe highest-leverage way to build with AI is to avoid heavily scaffolding or constraining the model into a fixed workflow (step one, then step two, then step three); giving the model a goal and a minimal set of tools and letting it decide how to use them produces better results than tightly orchestrating it.conf: high
• explicitScaffolding built to compensate for a current model's weaknesses typically improves performance by only about 10-20%, and that gain is usually wiped out by the next model generation, so the Claude Code team deliberately built for 'the model six months from now' from the start, accepting weak product-market fit in the interim rather than optimizing for the model that existed at the time.conf: high
  supportsStop Overengineering Your Agent Harness · p6 — Both describe the same mechanism — harness/scaffolding features expire because they encode an assumption about a model limitation that becomes false once the model improves — with Boris's account giving the quantified magnitude (10-20% typical gain) the other extraction's claim doesn't include.
• explicitdevelopedEvery person on the Claude Code team codes, including the product manager, engineering manager, designer, and finance lead; roles that used to be distinct (engineering, product management, design) now overlap by roughly half, and Boris expects the title 'software engineer' to eventually be replaced by something like 'builder.'conf: high
  supportsWhen AI Writes Almost All Code, What Happens To Software Engineering? · p16 — Both make the same claim — that product management and software engineering are converging rather than one replacing the other — with Boris's firsthand account of his own team (universal coding, ~50% role overlap) as independent, lived corroboration.
  extends ←This Week in AI: Chips, Checks, and Changing Jobs · p7 — Same person and team; this digest reports Cherny's more granular five-way functional decomposition (prototyper/builder/sweeper/grower/maintainer) of the general 'roles blur, everyone codes' claim made in the fuller interview already in this corpus.
• explicit'Latent demand' — building for what users are already improvising a product into doing — is 'the single most important principle in product,' and the same logic applies to models: watching what a model tries to do on its own and designing minimal scaffolding around that, rather than boxing the model into a predetermined role.conf: high
• explicitAnthropic deliberately releases agent products (Claude Code, then Cowork) early and imperfect specifically to study real-world safety behavior, because alignment research and sandboxed evaluation (the first two of three safety 'layers') can look good while a model still behaves differently once it is acting in the wild; the third layer can only be studied by shipping.conf: high
• explicitAnthropic frames open-sourcing safety-relevant infrastructure, such as Claude Code's agent sandbox (built to work with any agent, not just Claude Code), as participating in a deliberate 'race to the top' — competing with other labs/vendors to raise the industry's safety floor rather than to create proprietary advantage.conf: high
• explicitClaude Code's growth was not an immediate hit from launch: it took roughly eight to nine months (February to November) of modest adoption before crossing into '100% of my code' territory for Boris himself, with distinct inflection points tied to specific model releases (notably Opus 4 in May, and later models in November) rather than to product changes alone.conf: high
• explicitThe people most rewarded over the next few years will be curious, cross-disciplinary generalists who cross engineering, product, design, and business-sense boundaries — not narrow specialists who merely use AI tools well.conf: high
  supportsThe AI paradox: More automation, more humans, more work | Dan Shipper · p10 — Both claim that curious generalists outperform narrow specialists in the agentic-AI transition; this interview's advice to builders independently corroborates the other Lenny's Podcast episode's claim about who gains leverage.
• explicitBoris uses the printing press as the closest historical analogy for the current moment: literacy was below 1% of the population and gated by scribes before Gutenberg, and even after the printing press it took roughly 200 years to reach about 70% global literacy, which he uses to suggest the transition to near-universal programming ability will be similarly disruptive and uneven for a long time, even though he expects it to eventually happen.conf: high
• explicitAn informal Twitter poll found that about 70% of engineers and product managers report enjoying their jobs more since adopting AI tools (versus about 10% reporting less), while only 55% of designers report enjoying their jobs more (20% less) — a notably different distribution for designers that neither Lenny nor Boris can fully explain; Boris offers that Anthropic's own designers, who already code and were hired through technical screening even for non-technical roles, may be an atypical, non-generalizable sample.conf: high
• implicitBy running the same underlying Claude Code agent across every surface (terminal, desktop app, iOS, Slack, GitHub) and letting users split work across interfaces roughly a third each, Boris treats the choice of interface/form-factor as nearly irrelevant to the underlying work itself — implying that for Anthropic, competitive differentiation is expected to live in the model and its tool-use capability, not primarily in the surrounding product surface.conf: medium

Arguments

Narratives

• Boris, accustomed to debugging memory leaks himself via heap snapshots and a specialized debugger, watched a newer team member simply ask Claude Code to investigate a memory leak; Claude performed the same investigative steps Boris would have (took a heap snapshot, wrote a disposable just-in-time analysis tool, found the issue) and shipped a fix faster than Boris's own manual process.
  → Illustrates the 'mindset shift' cost Boris names explicitly — that even a deeply experienced practitioner defaults to old workflows and has to consciously 'transport himself to the current moment' — making this a concrete, personal instance of the skill/practice-obsolescence question running through the whole interview.
• A reported historical interview with a 15th-century scribe, asked how they felt about the printing press, who said they were relieved to be freed from the tedious work of copying text between books so they could focus on the parts they actually enjoyed (illustrating/drawing art, bookbinding).
  → Boris uses this found anecdote as a direct analogy for his own stated relief at no longer doing 'the tedious part' of coding (git plumbing, dependency wrangling) in favor of 'what to build' — explicitly mapping his own emotional experience onto a five-centuries-old precedent to normalize and legitimate his own practice change.
• A chain of 'latent demand' case studies: Facebook Marketplace emerging from the observation that ~40% of Facebook Groups posts were informal buying/selling; Facebook Dating emerging from the observation that 60% of profile views were between non-friends of the opposite gender; and Cowork emerging from watching non-engineers (a data scientist growing tomato plants, someone analyzing their genome, someone recovering wedding photos from a corrupted hard drive) improvise Claude Code via a bare terminal for non-coding tasks.
  → A sequence of analogous stories used to justify generalizing 'latent demand' from an established product-management heuristic into Boris's broader claim that the same logic applies to designing for models, not just for human users — the repetition across three unrelated products is the implicit argument for why the principle should be trusted.

Questions addressed

• ✅ Is coding, for an experienced practitioner like Boris, already a fully solved problem?
  Yes, for 'the kind of programming that I do' — but with an explicit caveat that full hands-off operation isn't yet safe at scale, since a human still needs to check correctness and safety, especially when many people are running the agent.
• ✅ What is the next frontier once coding itself is solved?
  Branching into general, non-coding computer-use tasks (Cowork) and models that proactively generate their own ideas and prioritization from feedback/telemetry, rather than only executing assigned coding tasks.
• ⭕ Why do designers, per Lenny's informal poll, report enjoying their jobs less since AI adoption at a notably higher rate (20%) than engineers or PMs (~10%)?
  Boris can only offer that Anthropic's own designers are an atypical, already-coding, technically-screened sample, and says he'd like to hear more from designers outside that context.
• ⭕ If junior engineers no longer accumulate the 'reps' of manually reviewing and being reviewed by human colleagues (because Claude now reviews 100% of PRs first), how do they develop the judgment that manual code review used to teach, and does Boris's own team — self-selected, mission-driven, pre-screened for coding aptitude even in non-technical roles — generalize to a team built from a normal hiring pool?

Gaps

• incompleteBoris states Claude reviews 100% of PRs and a human layer still exists 'after' that, but the interview never specifies what the human reviewer is actually checking for that Claude's own review doesn't already catch, nor how often the human layer actually changes or blocks something Claude approved. Without that detail, '100% AI-written, 100% AI-reviewed, plus a human layer' could describe either a meaningful independent check or a rubber stamp.
  Read next: Look for more granular accounts (internal engineering blog posts, case studies) of what Anthropic's human PR-review layer actually catches or overturns after Claude's automated review.
• incompleteThe 'underfund projects' and 'give unlimited tokens, don't optimize cost early' advice both describe practices inside Anthropic itself — an unusually well-capitalized, mission-selected, highly technical organization that screens for coding aptitude even in non-technical roles. The interview never engages whether either practice transfers to a resource-constrained team, or to a team whose non-engineers were not pre-selected for coding ability.
  Read next: Look for accounts of the same practices (deliberate team-size restraint, early token permissiveness) applied inside organizations without Anthropic's capital or hiring-bar advantages.
• incompleteBoris's firsthand account gives a concrete, affirmative answer to whether AI-driven role blurring favors smaller, cross-functional teams over specialized roles (yes, at the Claude Code team) — but neither Boris nor Lenny engages the two other sub-questions this corpus has separately flagged as open about the same shift: how career ladders change when junior engineers no longer accumulate manual code-review reps, and how interpersonal trust between human teammates is affected when most review traffic is now mediated by an agent rather than by colleagues reading each other's code.
  Read next: Look for accounts specifically addressing junior-engineer skill formation and team trust dynamics under agent-mediated review, as distinct from the team-structure question this interview does address.

Closes gaps elsewhere

• Agentic AI and Code Reviews · g3 — Gives a concrete, affirmative answer to one of the three sub-questions this gap names as unexplored: role blurring does favor smaller, cross-functional teams over specialized roles, as shown by the Claude Code team's own structure (every role codes, roughly 50% overlap between engineering/product/design, with Boris predicting the title 'software engineer' itself may disappear). The gap's other two sub-questions — career-ladder effects on junior engineers and interpersonal trust effects of agent-mediated review — remain unaddressed by this interview and are carried forward as this extraction's own gap.

Related documents

• Stop Overengineering Your Agent Harness
• Designing the AI-native engineering organization
• Agentic Code Review
  Code Was Never the Bottleneck Human-in-the-Loop Pull Requests & Async Code Review
• The AI paradox: More automation, more humans, more work | Dan Shipper
  Code Was Never the Bottleneck Product Engineer (PM/Engineering Role Merge) Riding Model Improvements
• AI isn't making developers more productive – it's making them busier
  Code Was Never the Bottleneck Pull Requests & Async Code Review
• The "funhouse mirror": How AI reflects the hidden truths of your software pipeline
  Code Was Never the Bottleneck Human-in-the-Loop
• Anthropic Details How It Contains Claude Across Web, Code, and Cowork
  Claude Code Human-in-the-Loop
• Automated Bug Fixing at Facebook Scale
  Human-in-the-Loop Pull Requests & Async Code Review
• How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
  Claude Code Riding Model Improvements
• Inside Claude Code Auto Mode: Anthropic's Autonomous Coding System with Human Approval Gates
  Claude Code Human-in-the-Loop