When Boris Cherny, the creator and head of Claude Code at Anthropic, shared his personal terminal workflow recently, the software engineering community took notice — and quickly recognized a transformative shift in AI-driven software development.
Over the past week, developers and industry insiders have dissected Cherny’s detailed explanation of how he leverages multiple AI agents simultaneously, enabling a single programmer to match the output traditionally requiring an entire engineering team. This novel approach has sparked widespread discussion about the future of coding and AI’s role in it.
Parallel AI Agents: Turning Coding into a Real-Time Strategy
Unlike conventional software development, which follows a linear process of writing, testing, and iterating code, Cherny operates more like a commander overseeing multiple AI “units.” He runs five instances of Claude in parallel within his terminal, each dedicated to distinct tasks such as running test suites, refactoring legacy code, or drafting documentation. System notifications alert him whenever an AI agent requires input, allowing seamless multitasking.
Additionally, Cherny manages five to ten instances of Claude via a browser interface, utilizing a “teleport” command to transfer sessions between the web and his local environment. This method exemplifies Anthropic’s “do more with less” philosophy, focusing on orchestrating existing AI models effectively rather than expanding infrastructure exponentially.
The Strategic Choice of a Slower, Smarter AI Model
In a surprising departure from the tech industry’s emphasis on speed, Cherny exclusively uses Anthropic’s largest and slowest AI model, Opus 4.5, for all coding tasks. Although slower, this model requires less human intervention and offers superior tool usage, ultimately accelerating overall productivity by reducing the time spent correcting errors.
This insight is crucial for enterprise leaders evaluating AI investments: the real bottleneck lies not in processing speed but in minimizing human correction efforts. Paying upfront for a more capable model can save significant time and resources downstream.
Creating a Self-Correcting AI Through Shared Knowledge
To overcome the common limitation of AI “amnesia” — the inability to retain knowledge across sessions — Cherny’s team maintains a single shared file, CLAUDE.md, within their git repository. Whenever Claude makes a mistake, it is documented in this file, effectively teaching the AI to avoid repeating errors in future sessions.
This approach transforms the codebase into a living system that continuously improves. Human reviewers update the AI’s instructions alongside code fixes, turning every error into a new rule. The longer the collaboration, the smarter and more reliable the AI becomes.
Automation Through Slash Commands and Specialized Subagents
Cherny enhances efficiency further by employing custom slash commands embedded within the project repository. One notable example is the /commit-push-pr command, which automates the entire process of committing code, pushing changes, and opening pull requests, eliminating manual steps.
Specialized subagents handle discrete development phases, such as simplifying code architecture or running comprehensive end-to-end tests. This layered automation offloads tedious tasks from human developers, allowing them to focus on higher-level problem-solving.
Verification Loops: Ensuring AI-Generated Code Quality
A key factor behind Claude Code’s reported rapid growth to $1 billion in annual recurring revenue is its built-in verification loop. Cherny highlighted that Claude tests every change it implements through browser automation and test suites, iterating until the code works correctly and the user experience meets standards.
By enabling the AI to verify its own output, the workflow substantially improves code quality and reliability, multiplying effectiveness by two to three times compared to unverified AI-generated code.
Implications for the Future of Software Engineering
Cherny’s workflow signals a fundamental shift in how programming is conducted. AI is no longer merely an autocomplete tool but an integrated labor force that multiplies human developers’ productivity severalfold.
As Jeff Tang, a respected developer, summarized on social media: “Read this if you’re already an engineer… and want more power.” The key to unlocking this potential is embracing AI as a workforce rather than just an assistant.
Developers who adopt this mindset early will not only boost their output but change the very nature of software creation — leaving traditional typing-based coding far behind.
Related Articles
Jeff Bezos’ Prometheus Raises $12 Billion at $41 Billion Valuation Amid AI Race
Gridcare Secures $13.3 Million to Uncover Hidden 100 GW Data Center Capacity in Electrical Grid
Google Pledges to Expand AI Infrastructure Capacity by 1000x Within Five Years
Is Google DeepMind turning Gemini into the agent platform ChatGPT still has to answer?
