Boris Cherny: Claude Code Ups Anthropic Productivity by 150%

Anthropic's development strategy for Claude Code centers on building for the model six months in the future, not the current one. This approach acknowledges the rapid pace of LLM advancements, where today's limitations quickly become tomorrow's capabilities. Founders building on LLMs are advised to anticipate future model strengths rather than optimizing for present weaknesses. This forward-looking strategy means Claude Code undergoes continuous and extensive rewriting. There is no part of the current Claude Code that existed six months ago, highlighting the constant adaptation required. This iterative process involves trying new features, gathering user feedback, and learning from usage patterns.

Claude Code is in a state of constant flux, with its entire codebase being rewritten and updated frequently. Tools are unhipped and new ones are added every couple of weeks, meaning no part of Claude Code from six months ago remains in the current version. Most of the current Claude Code base, approximately 80%, is less than a couple of months old. This short shelf life for code is becoming the norm, especially for leading founders and innovative companies. This continuous development cycle highlights the dynamic nature of AI-driven software development. Anthropic has experienced significant productivity gains since implementing Claude Code. While the team doubled in size last year, productivity per engineer grew by approximately 70%. Since Claude Code's release, this figure has surged to a 150% increase in productivity per engineer, measured by pull requests and cross-checked against commits and commit lifetimes. This level of productivity increase is unprecedented in the industry.

Claude Code's initial terminal-based interface was an accidental outcome, chosen for its ease of prototyping. Boris Cherny, the creator, initially built a simple terminal chat app to understand Anthropic's API, not intending it to be the final product. This minimalist approach allowed for rapid iteration without the overhead of building a complex UI. The terminal's longevity and effectiveness surprised its creators, who initially expected it to have a short lifespan. Despite its limitations (e.g., 80x100 characters, 256 colors, no mouse interactions), the CLI proved highly functional and enjoyable for developers. This simple form factor enabled a focus on core utility and user experience.

A core principle guiding Claude Code's development is 'latent demand' – building solutions for tasks users are already trying to accomplish. Features like 'plan mode' emerged directly from observing users attempting to structure their AI interactions, asking Claude to plan out ideas without immediately writing code. This insight led to the rapid implementation of plan mode, which was developed in just 30 minutes based on user feedback. This approach contrasts with creating entirely new user behaviors; instead, it focuses on making existing, often cumbersome, workflows easier. The team actively monitors GitHub issues, internal Slack channels, and even observes users directly to identify these needs.

The rapid evolution of AI models necessitates a 'beginner's mindset' and humility from engineers. Traditional engineering values, such as strong opinions and extensive experience, can become liabilities when model capabilities change so quickly. The most effective engineers in this new paradigm are those who can think scientifically and from first principles, constantly questioning assumptions. This shift impacts hiring, with an emphasis on candidates who can recognize and learn from their mistakes. The ability to adapt and embrace new ways of working, rather than clinging to outdated methods, is crucial. This adaptability is exemplified by engineers who leverage AI to automate tasks they previously performed manually.

Anthropic operates under the principle of 'the bitter lesson': never bet against the model. This means that given enough time, a more general AI model will always outperform a more specific, hand-engineered solution. This philosophy influences decisions about building scaffolding (code that enhances the model's performance) versus waiting for the next model release. While scaffolding can offer a temporary performance boost (e.g., 10-20%), these gains are often wiped out by the next model iteration. Consequently, the team frequently unships tools and deletes scaffolding as model capabilities improve. This constant re-evaluation ensures that engineering effort is focused on areas where it provides lasting value.

The vision for Claude Code extends to collaborative AI agents and 'agent topologies.' This involves configuring multiple agents with fresh, uncorrelated context windows to tackle complex problems. This approach, exemplified by the recent launch of Claude Teams, allows agents to communicate and build larger, more sophisticated projects. An early success story is the development of Claude Code's plugins feature, which was entirely built by a 'swarm' of agents over a weekend with minimal human intervention. This involved agents creating tasks on a Trello board, spawning sub-agents to pick up tasks, and independently figuring out solutions. This demonstrates the potential for AI to automate entire development processes.

Boris Cherny was drawn to Anthropic after experiencing the "breathtaking" capabilities of early AI products. He connected with Ben Mann, one of Anthropic's founders, and was won over by the team's approach. Two key factors influenced his decision: Anthropic's operation as a research lab focused on building safe models, where the model itself is paramount over product, and its deeply mission-driven culture. As a sci-fi reader, Cherny is acutely aware of the potential negative outcomes of advanced AI. He sought a workplace that internalized the importance of AI safety, a topic frequently discussed among employees at Anthropic. This commitment to safety resonated deeply with his personal values.

Boris Cherny predicts that coding will be "generally solved for everyone" in the near future, similar to how it is already 100% solved for him personally since Opus 4.5. He has uninstalled his IDE and lands around 20 pull requests daily, entirely generated by Claude Code. Across Anthropic, 70-90% of code is written by Claude Code, with many teams and individuals reaching 100%. This trend suggests the title "software engineer" may eventually disappear, replaced by roles like "builder" or "product manager." Engineers are becoming generalists, with PMs, designers, and even finance personnel on Cherny's team now coding. The upper bound of this trend involves reaching ASL4 (AI Safety Level 4), which is a significant milestone in AI development.

The development of Co-work was driven by overwhelming demand from non-technical users who were already adopting Claude Code for diverse tasks, such as monitoring tomato plants, recovering wedding photos, and financial analysis. Recognizing this need, Anthropic developed Co-work as a user-friendly wrapper for Claude Code. Co-work is essentially Claude Code running under the hood, utilizing the same agent. It was developed rapidly, with Felix and his team building it in approximately 10 days, entirely written by Claude Code. Designed for non-technical audiences, Co-work includes crucial safety features like running all code in a virtual machine, robust deletion protections, and extensive permission controls.