Overview
The engineering community is dissecting a workflow that fundamentally redefines software development, originating from a simple terminal thread posted by Boris Cherny, head of Claude Code at Anthropic. What started as a personal setup guide has become a viral manifesto, demonstrating how advanced AI agents can elevate a single developer to the output capacity of a small, specialized team. The core revelation is that modern coding is no longer a linear process of writing, testing, and moving on; it is a command-and-control operation.
Cherny’s method shifts the developer's role from syntax writer to fleet commander. By running multiple AI agents simultaneously, the process transforms into something akin to a real-time strategy game, where the human orchestrates autonomous units rather than typing individual lines of code. This approach challenges the industry assumption that faster token generation equals superior development velocity.
This detailed look at the workflow reveals a powerful strategy: maximizing output by managing parallel workstreams and minimizing the costly human time spent correcting AI errors.
The Parallel Command Structure of Modern Coding
The Parallel Command Structure of Modern Coding
Cherny’s most striking contribution is the dismantling of the traditional "inner loop" of development. In conventional programming, a developer completes a function, tests it, and then proceeds sequentially to the next module. Cherny’s setup bypasses this bottleneck by running five or more distinct Claude agents concurrently within a single terminal environment.
The execution relies on meticulous system management, utilizing tools like iTerm2 system notifications to manage five simultaneous, specialized workstreams. While one agent is dedicated to running a comprehensive test suite, a second agent might be refactoring a deeply embedded legacy module, and a third might be drafting technical documentation for the same feature. These agents are not merely running tasks; they are operating in parallel, simulating the specialized roles of multiple human engineers.
Furthermore, the workflow extends beyond the local terminal. Cherny integrates web-based AI sessions, running five to ten instances of Claude directly in the browser. The use of a "teleport" command allows him to hand off context and tasks between the web environment and his local machine. This orchestration capability validates a broader strategic thesis: superior system design and agent coordination can yield exponential productivity gains, even when compared to competitors focused solely on massive infrastructure build-outs.
Choosing Intelligence Over Speed The Opus 4.5 Insight
In a sector obsessed with minimizing latency and maximizing token generation speed, Cherny made a counterintuitive declaration: he exclusively uses Anthropic's heaviest, slowest model, Opus 4.5, for all coding tasks. This choice represents a critical, high-level insight for enterprise technology leaders.
The argument centers on the distinction between the "compute tax" and the "correction tax." While Sonnet or other smaller, faster models generate tokens quickly, they require significantly more human intervention—the "correction tax"—to fix architectural flaws, logical errors, or style inconsistencies. Cherny argues that paying the higher computational cost for the smartest model upfront, which is better at tool use and complex reasoning, ultimately saves time and effort later.
This major change reframes the bottleneck in AI development. The limiting factor is not the speed at which the model generates the next token; it is the human cognitive time required to validate, debug, and correct the model's inevitable mistakes. By choosing the most capable model, the developer minimizes the time spent in the iterative loop of error correction, making the slower, smarter model the fastest option overall.
Building a Self-Correcting Codebase
Beyond managing parallel tasks and selecting the optimal model, Cherny’s team has solved one of the most persistent and frustrating problems in large-scale AI development: AI amnesia. Standard large language models operate in isolated sessions and do not inherently retain a company's specific architectural decisions, coding conventions, or historical mistakes from one interaction to the next.
To combat this, the team maintains a single, dedicated file named CLAUDE.md within their core git repository. This file functions as a living, collective memory for the AI. Whenever an agent produces an incorrect output or deviates from established best practices, the mistake is documented and added to CLAUDE.md. This practice systematically transforms the codebase itself into a self-correcting, continuously learning entity.
This mechanism moves the relationship between the developer and the AI from a simple prompt-response cycle to a structured, iterative feedback loop. It ensures that the accumulated institutional knowledge of the company—its unique style guide, its historical debt, its specific failure points—is constantly reinforced and available to the AI agents, dramatically increasing the reliability and coherence of the output.
