Claude Code

Claude Code is Anthropic’s coding agent: a terminal-first AI tool that writes, reviews, and executes code by using tools autonomously. Created by Boris Cherny as a solo hack in 2024; released externally in February 2025; by 2026 accounts for approximately 4% of all public GitHub commits.

Origins and design philosophy

Claude Code grew out of Anthropic’s internal model development roadmap: coding → tool use → computer use. This sequence was chosen deliberately — coding is the domain where agentic capability can be studied most rigorously, and where safety can be verified most precisely.

The terminal interface was not a deliberate design choice. It was the fastest way to build alone. The team later considered other form factors but found the terminal was the only interface that could keep pace with rapid model improvement.

The first internal demo received two likes. External release in February 2025 was not immediately a hit; adoption grew slowly until the first sharp inflection with Opus 4 / Sonnet 4 in May 2025.

Core design inversion

Most LLM products treat the model as a component inside a larger system with pre-specified tool sequences. Claude Code inverted this:

‘The product is the model.’

Minimal scaffolding. Minimal pre-specified tool order. The model receives tools and a goal, and decides how to use them. This is the product expression of Latent Demand form 2 (model latent demand) and Bitter Lesson — scaffolding buys at most 10–20% improvement and is typically wiped out by the next model.

Growth and impact

4% of all public GitHub commits attributed to Claude Code (Semi-Analysis, 2026) [?source]
Private repository usage estimated meaningfully higher
Daily active users have continued to accelerate (doubled in the month before the Feb 2026 interview)
At Anthropic: ~200% productivity increase (measured in PRs) despite ~4× engineering headcount growth

Features and form factors

Available across: terminal (all platforms), desktop app (code tab), iOS/Android, Slack integration, GitHub integration, IDE extensions, web.

Key features:

Plan mode — injects ‘please don’t write any code yet’ into the prompt; model plans in dialogue before executing.
Auto-accept edits — after plan approval, the model executes without confirmation checkpoints.
Multi-Claude-ing — running multiple sessions in parallel across different tasks simultaneously.
Claude-reviewed PRs — at Anthropic, Claude reviews 100% of pull requests; human review follows.

Cowork

Claude Code’s underlying agent also powers Cowork — Anthropic’s desktop-native product for non-engineering tasks (email, project management, web forms, calendar). Cowork was built in 10 days using Claude Code itself. It ships with a sandboxed VM for safety and a Chrome integration for browser-based automation.

The same Claude Code agent runs in Cowork; the difference is packaging and the target user (non-engineers).

Safety model

Claude Code was used internally at Anthropic for 4–5 months before external release — specifically to study agentic safety in the wild, since it was the first coding agent to be broadly deployed. Anthropic open-sourced its agent sandbox (compatible with any agent) as part of a ‘race to the top’ strategy.

The three-layer safety model Boris describes:

Alignment + Mechanistic Interpretability (training-time)
Evals (laboratory)
In-the-wild behaviour (deployment)

Practical tips (Boris Cherny’s own usage)

Use Opus 4.6 with max effort — usually not more expensive due to fewer correction loops.
Start in plan mode (~80% of tasks); auto-accept after plan is agreed.
Try all form factors; the same agent runs everywhere.
Run multiple Claudes in parallel.

Mike Krieger: Anthropic as patient zero

Mike Krieger on Anthropic, AI Timelines, and Product in a 90% AI-Written Codebase provides the CPO’s view of what Claude Code has done to Anthropic’s own engineering culture.

The numbers: over 90% of the code in the company is written by AI. The Claude Code team is estimated at 95%+ — the tool is substantially built by itself. Krieger calls Anthropic ‘patient zero’ for AI-native software development.

The new bottlenecks: when engineering velocity is no longer the constraint, the critical path moves to two new places:

Upstream — deciding what to build. Alignment on priorities becomes the load-bearing work.
Downstream — the merge queue. Individual engineers (human or AI) produce changes faster than the collective deployment pipeline can absorb coherently. The constraint moves from writing code to getting code into production without regressions.

Code review has changed: the model is acceptance testing rather than line-by-line review. Does this change do what it should do? Claude reviews 100% of PRs at Anthropic; human review follows.

LLM-friendly codebase: the Anthropic codebase is designed with the constraint that Claude should be able to read and edit it. Clearly named functions, consistent patterns, minimal magic. This is a deliberate design choice, analogous to mobile-first design as a constraint for the smartphone era.

Product × research embedding: the products that are genuinely differentiated (Artifacts, Projects) emerged from the intersection of post-training decisions and product design, not from UX work on top of a fixed model. Krieger identifies the PM role inside model training and post-training work as the highest-leverage position at a model company.

Zevi Arnovitz: Claude Code for non-technical PMs

Zevi Arnovitz on Vibe Coding as a PM, Multi-Model Review, and Building with Claude Code provides the clearest practitioner account of Claude Code used by someone with zero technical background. Zevi (PM at Meta) built and ships a paying product — StudyMate — using Claude Code running inside Cursor. His /commands workflow uses CLAUDE.md as the team memory and encodes each development phase as an invokable prompt. He runs Claude Code as primary dev lead alongside Codex and Cursor Composer for multi-model peer review. See Vibe Coding for the full account.