What coding's automation means for software engineering and beyond

Executive overview

Coding is largely solved. AI now writes 100% of code for engineers like Boris Cherny, head of Claude Code at Anthropic, with productivity per engineer up 200%. The shift from writing code to directing AI agents changes not just engineering but every computer-based role.

The next frontier is AI that proposes ideas, not just executes them — reading feedback, scanning telemetry, and surfacing what to build next. Roles like product manager, designer, and data scientist are next in line.

The leverage is not in writing less code — it's in being freed to think about what to build.

The origins and growth of Claude Code

• Started as a solo hack in a terminal; got two internal likes when first announced.
• Terminal form factor chosen because it was the only way to keep pace with rapidly improving models.
• 4% of all public GitHub commits are now authored by Claude Code; private repo share is likely higher.
• Growth rate is still accelerating — daily active users doubled in the past month.
• Not an immediate hit on external launch; took months for users to understand what it was.

Why 100% AI-generated code is now normal

• Boris has not edited a single line of code by hand since November.
• Ships 10–30 pull requests per day, with five agents running in parallel during this recording.
• Claude also reviews 100% of pull requests at Anthropic, with a human layer still on top.
• Productivity gains of hundreds of percent are unprecedented — previous industry norms were a few percentage points per year from dedicated tooling work.
• New team members approach problems more "AGI-forward" than engineers who learned older habits; the memory-leak example shows a newer engineer simply asked Claude to diagnose it, faster than manual heap analysis.

What comes after coding

• Coding is "largely solved" for the kinds of programming Boris does day-to-day.
• The next shift: Claude proactively proposes ideas by scanning Slack threads, bug reports, and telemetry.
• Agentic AI — LLMs that can act on tools, not just converse — is coming to non-technical roles for the first time via products like Claude's computer-use mode.
• Adjacent roles to watch: product managers, designers, data scientists, anyone whose work lives on a computer.

Latent demand as a product principle

• Latent demand: build where people already are; watch how they misuse existing products to discover what to build next.
• Facebook Marketplace emerged from 40% of Facebook group posts being buy/sell activity.
• Claude's computer-use product emerged from seeing people use a coding terminal to analyze MRIs, recover wedding photos, and manage genome data.
• Modern extension: watch what the model tries to do, and reduce friction around it — invert the box, give the model tools and let it decide.

Team structure and the "underfunding" principle

• Underfunding forces teams to automate — engineers given fewer resources ship faster by leaning harder on Claude.
• Everyone on the Claude Code team codes: PM, EM, designer, finance, data science.
• Give engineers unlimited tokens early; optimize token cost only once the idea is validated.
• The role boundaries between engineering, PM, and design are blurring; "builder" may replace "software engineer" as a title by end of year.

Principles for building AI products

• Don't box the model in with rigid step-by-step orchestration; give it tools and a goal.
• The bitter lesson: the more general model always wins long-term; scaffolding gains of 10–20% are often erased by the next model release.
• Build for the model six months from now, not the model today; product-market fit will be weak at first, then click when the better model ships.
• Bet on models improving at: long autonomous runs (from 30 seconds to 30+ minutes unattended), tool use, and computer use.

Tips for getting the most from Claude Code

• Use the most capable model (currently Opus 4.6) — it often uses fewer tokens than cheaper models because it requires less correction.
• Start tasks in plan mode (shift-tab twice in terminal) before letting Claude write code; auto-accept edits once the plan looks good.
• Run multiple agents in parallel — terminal, desktop app, iOS, and Slack integration all run the same underlying agent.
• No single right way to use it; ask Claude Code itself for setup recommendations.

Safety, interpretability, and releasing early

• Three layers of safety: mechanistic interpretability (neuron-level model monitoring), evals (lab Petri-dish testing), and real-world behavior.
• Claude Code was used internally at Anthropic for four to five months before external release specifically to study agent safety.
• Superposition: in large models, a single neuron encodes dozens of concepts simultaneously; understanding this is an active research frontier.
• Early releases are a safety valve — real-world behavior is the only way to validate alignment beyond the lab.

The broader shift: printing press analogy

• Pre-printing-press Europe: sub-1% literacy; scribes held all knowledge work.
• 50 years after Gutenberg: more printed material than in the prior 1,000 years; cost dropped ~100x.
• Literacy still took 200 years to reach 70% globally — transitions are fast at the frontier, slow in diffusion.
• Programming going from a specialist skill to a universal one follows the same arc; the Renaissance required mass literacy; the next leap requires mass programmability.
• Near-term it will be disruptive and painful; long-term it is democratizing.

Advice for staying ahead

• Experiment with AI tools now; stay on the bleeding edge.
• Become a generalist — cross disciplines (engineering + design sense, engineering + business judgment, PM + coding).
• The people rewarded most will be curious generalists who can hold the full problem, not just their slice.