Boris Cherny’s Workflow: How Anthropic’s Claude Code Creator is Redefining Software Engineering with AI Agents

When the creator of the world’s most advanced coding agent speaks, Silicon Valley doesn’t just listen – it takes notes. The insights from such a figure, particularly one at the forefront of artificial intelligence, carry significant weight, often signaling shifts in industry paradigms. For the past week, the engineering community across the globe has been meticulously dissecting a thread on X from Boris Cherny, the visionary creator and head of Claude Code at Anthropic. What began as a seemingly casual sharing of his personal terminal setup has rapidly escalated into a viral manifesto on the future of software development, with numerous industry insiders hailing it as a watershed moment for the burgeoning AI startup, Anthropic.

The profound impact of Cherny’s revelations quickly resonated through the developer community. "If you’re not reading the Claude Code best practices straight from its creator, you’re behind as a programmer," asserted Jeff Tang, a prominent voice and respected figure in the developer community, emphasizing the immediate relevance of Cherny’s insights for practitioners. Kyle McNease, another influential industry observer, took his assessment even further, declaring that with Cherny’s "game-changing updates" and the implications of his workflow, Anthropic is "on fire," potentially experiencing "their ChatGPT moment." This comparison highlights the potential for Anthropic to achieve a similar level of breakthrough recognition and market disruption that OpenAI experienced with the launch of ChatGPT.

The palpable excitement surrounding Cherny’s workflow stems from a fascinating paradox: his operational approach is remarkably simple in its underlying principles, yet it empowers a single human developer to operate with the output capacity traditionally associated with a small, dedicated engineering department. This exponential leap in productivity has captivated observers. As one user noted on X after successfully implementing Cherny’s setup, the entire coding experience "feels more like Starcraft" than traditional coding. This powerful analogy vividly illustrates a fundamental shift from the laborious, line-by-line process of typing syntax to the strategic oversight of commanding autonomous units – in this case, sophisticated AI agents – to execute complex tasks. This article delves into an in-depth analysis of the revolutionary workflow that is rapidly reshaping how software is built, drawing directly from the architect himself.

How Running Five AI Agents at Once Transforms Coding into a Real-Time Strategy Game

Perhaps the most striking and immediately impactful revelation from Cherny’s disclosure is his departure from the conventional, linear approach to software development. In the traditional "inner loop" of development, a programmer typically focuses on a single task: writing a function, meticulously testing it, debugging, and only then moving on to the next isolated task. This method, while structured, can often be bottlenecked by human sequential processing. Cherny, however, adopts a fundamentally different role, positioning himself not as a coder, but as a "fleet commander" orchestrating multiple simultaneous operations.

"I run 5 Claudes in parallel in my terminal," Cherny explicitly stated, outlining the core of his multitasking strategy. "I number my tabs 1-5, and use system notifications to know when a Claude needs input." This setup allows him to leverage the power of concurrent processing by distributing different development tasks across multiple AI agents. By utilizing iTerm2 system notifications – a feature that provides alerts when a specific terminal tab requires attention – Cherny effectively manages these five simultaneous work streams. While one AI agent might be diligently running a comprehensive test suite to ensure code quality, another could be engaged in the intricate process of refactoring a legacy module, and a third might be drafting detailed documentation for newly developed features. This parallel execution dramatically reduces idle time and accelerates project timelines.

Beyond his local terminal, Cherny also maintains "5-10 Claudes on claude.ai" in his browser, demonstrating a seamless integration between local and cloud-based AI interactions. He employs a clever "teleport" command, which allows him to hand off active sessions between the web interface and his local machine, ensuring continuity and flexibility in his workflow. This sophisticated orchestration of AI agents provides a tangible validation of the "do more with less" strategy recently articulated by Anthropic President Daniela Amodei. While competitors, notably OpenAI, are actively pursuing massive, trillion-dollar infrastructure build-outs to scale their AI capabilities, Anthropic is demonstrating that superior strategic orchestration and intelligent deployment of existing, powerful models can yield exponential productivity gains, effectively multiplying human output without requiring proportionate increases in raw compute power.

The Counterintuitive Case for Choosing the Slowest, Smartest Model

In an industry perpetually obsessed with latency and the sheer speed of computational output, Cherny’s choice of AI model is particularly surprising and, for many, counterintuitive. He revealed that he exclusively uses Anthropic’s heaviest and, consequently, slowest model: Opus 4.5. This decision flies in the face of the common perception that faster models inherently lead to greater efficiency.

"I use Opus 4.5 with thinking for everything," Cherny explained. "It’s the best coding model I’ve ever used, and even though it’s bigger & slower than Sonnet, since you have to steer it less and it’s better at tool use, it is almost always faster than using a smaller model in the end." This statement offers a critical insight for enterprise technology leaders and developers alike. Cherny argues that the true bottleneck in modern AI-assisted development is not the raw generation speed of the AI’s tokens, but rather the significant human time spent correcting the AI’s mistakes or guiding it through complex logic. Cherny’s workflow suggests a powerful economic principle: paying a higher "compute tax" for a more intelligent, capable model upfront, even if it’s slower in terms of initial output, ultimately eliminates a far greater "correction tax" later on. The superior accuracy and understanding of Opus 4.5 mean fewer iterations, less human intervention, and a more direct path to a correct and functional solution, thereby accelerating the overall development cycle.

One Shared File Turns Every AI Mistake into a Permanent Lesson

Another pervasive challenge in leveraging large language models (LLMs) for complex development tasks is the issue of "AI amnesia." Standard LLMs typically lack a persistent memory of a company’s specific coding style, intricate architectural decisions, or preferred best practices from one interaction session to the next. Each new prompt often starts from a blank slate, leading to repetitive corrections and inconsistencies.

To ingeniously address this, Cherny’s team maintains a single, crucial file named CLAUDE.md within their git repository. This Markdown file serves as a collective, evolving instruction set for their AI agents. "Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time," he wrote. This simple yet profound practice transforms the entire codebase into a self-correcting, continuously learning organism. When a human developer reviews a pull request and identifies an error or a deviation from established guidelines, they don’t just fix the code; they proactively tag the AI by updating its instructions within CLAUDE.md. As Aakash Gupta, a prominent product leader analyzing the thread, aptly noted, "Every mistake becomes a rule." The longer the team collaborates using this system, the more refined and intelligent their AI agents become, progressively aligning their output with the team’s specific requirements and institutional knowledge.

Slash Commands and Subagents Automate the Most Tedious Parts of Development

The "vanilla" workflow, as one observer admiringly described it, that Cherny employs is profoundly powered by the rigorous automation of repetitive and often tedious development tasks. Cherny makes extensive use of slash commands – custom shortcuts that are explicitly checked into the project’s repository – to handle complex operations with a single, efficient keystroke. This approach standardizes automation and makes it readily available to the entire team.

He highlighted a particularly impactful command, /commit-push-pr, which he invokes dozens of times daily. Instead of manually typing a series of git commands, crafting a detailed commit message, and then navigating to a platform to open a pull request, the AI agent autonomously handles this entire bureaucratic sequence of version control. This significantly streamlines the developer’s interaction with the repository.

Furthermore, Cherny strategically deploys specialized subagents – essentially AI personas configured with specific expertise and objectives – to manage distinct phases of the development lifecycle. For instance, he utilizes a dedicated "code-simplifier" subagent to meticulously clean up and optimize the architectural structure of the code after the primary development work is completed. Another specialized agent, a "verify-app" agent, is tasked with running comprehensive end-to-end tests to ensure the application’s integrity and functionality before any code is shipped to production. These subagents act as expert assistants, each taking on specific responsibilities to ensure quality and efficiency.

Why Verification Loops are the Real Unlock for AI-Generated Code

If there is a singular, overarching reason why Claude Code has reportedly achieved the significant milestone of $1 billion in annual recurring revenue so rapidly, it is likely the groundbreaking implementation of robust verification loops. This feature fundamentally transforms the role of AI in coding: the AI is no longer merely a text generator but evolves into a proactive, intelligent tester and quality assurance agent.

"Claude tests every single change I land to claude.ai/code using the Claude Chrome extension," Cherny revealed, detailing a powerful closed-loop system. "It opens a browser, tests the UI, and iterates until the code works and the UX feels good." This comprehensive approach means that the AI doesn’t just produce code; it actively validates its own work against real-world conditions. Cherny emphatically argues that empowering the AI with the capability to verify its own output – whether through sophisticated browser automation, executing bash commands within the terminal, or running extensive test suites – dramatically improves the quality of the final result by a factor of "2-3x." The profound implication is that the agent doesn’t just write code; it definitively proves that the code works as intended, significantly reducing the burden on human developers for initial verification and debugging. This self-verification capability, combined with other strategic investments like Anthropic’s acquisition of Bun, underscores the company’s commitment to building a holistic and highly effective developer ecosystem around Claude Code.

What Cherny’s Workflow Signals About the Future of Software Engineering

The overwhelming reaction to Boris Cherny’s detailed thread strongly suggests a pivotal and irreversible shift in how developers conceptualize and execute their craft. For many years, the concept of "AI coding" primarily evoked images of an autocomplete function within a text editor – a helpful, albeit limited, tool designed to speed up typing and offer minor code suggestions. Cherny, however, has powerfully demonstrated that AI can now function as an entire operating system for labor itself, transforming the programmer’s role from a hands-on builder to a strategic manager.

"Read this if you’re already an engineer… and want more power," Jeff Tang succinctly summarized on X, encapsulating the sentiment of many who see Cherny’s approach as a pathway to unprecedented individual productivity. The tools required to multiply human output by a factor of five, or even more, are no longer theoretical; they are demonstrably here and accessible. They demand, however, a fundamental shift in mindset: a willingness to stop thinking of AI merely as an assistant that helps with small tasks and to start treating it as a dynamic, autonomous workforce capable of executing complex, parallel operations. The programmers who are agile enough to make this crucial mental leap first will not only become significantly more productive but will also find themselves playing an entirely different game of software development. While they command their AI fleets, everyone else will still be typing, constrained by the limitations of traditional, linear coding practices.