A recent X (formerly Twitter) post by Boris Cherny, the mind behind Anthropic's Claude Code, has sent ripples through the developer community. Instead of announcing a new feature or a funding round, Cherny simply shared his daily workflow using the tool. What followed was an intense discussion, dissecting and debating his approach to AI-powered development.
Five Parallel Terminals, One Human Conductor
Traditional coding often feels like a linear process: write a function, run tests, fix bugs, then move to the next. Cherny's method flips this on its head. He simultaneously runs five Claude instances in his terminal, labeling them 1 through 5. He relies on iTerm2 system notifications to alert him when an agent requires input, allowing him to switch contexts seamlessly. One agent might be running a test suite, another refactoring an older module, and a third drafting documentation. In this setup, Cherny acts less like a direct coder and more like a strategic dispatcher.
The setup extends beyond the terminal. Cherny also maintains 5 to 10 claude.ai browser sessions, using a custom “teleport” command to fluidly transition these web sessions to his local terminal. This means his entire 'workbench' spans multiple environments, with numerous tasks progressing concurrently.
Why This Workflow Blew Up the Developer Sphere
This 'multi-front' approach is a significant departure from how most developers typically work. One commenter on X aptly noted that trying it felt 'more like playing Starcraft than writing code.' This analogy perfectly captures the shift in role: you're no longer typing line by line, but commanding a squad of autonomous units to execute tasks. This sentiment resonated widely, becoming a frequently cited description of the experience.
Industry figures quickly weighed in. Jeff Tang, for example, declared on X that any programmer not familiar with Cherny's Claude Code best practices was already falling behind. Kyle McNease, another observer, went even further, calling this workflow reveal Anthropic's 'ChatGPT moment' – a pivotal demonstration of AI's potential to redefine interaction paradigms.
Practical Implications for Developers
The true significance of this workflow isn't just the 'five agents' number, but its potential to reshape an individual developer's productivity structure. Historically, parallel processing of multiple modules was limited by the cognitive load of context switching. Now, each AI agent can independently manage a task stream, with the human only intervening at critical decision points. This aligns with Anthropic President Daniela Amodei's earlier emphasis on enabling 'fewer people to do more things.'
This model is particularly appealing for independent developers and small teams. It suggests the possibility of advancing feature development, testing, and documentation concurrently, even without a large engineering team. Of course, this also demands stronger task decomposition and prioritization skills from the developer – it's easy to assign tasks to five agents, but knowing which tasks are suitable for parallel execution and when to consolidate efforts is the real challenge.
A Reality Check: Workflow, Not a Silver Bullet
It's crucial to remember that Cherny hasn't claimed this method is universally applicable. He's showcasing a personal preference, built on a deep understanding of Claude's behavior. For the average developer, directly replicating a 5-terminal setup with notifications might not be the ideal starting point. A more pragmatic approach would be to begin with two parallel tasks and gauge your ability to manage the rhythm.
Furthermore, multiple agents running in parallel can lead to context fragmentation. Each agent operates in its own silo, unaware of what the others are doing. The coordination burden still falls on the human. While Cherny's system addresses 'when do I need to intervene' with notifications, it doesn't solve 'how do agents understand each other.' Official technical details on this specific aspect remain limited.
- If you're keen to try, start with two terminals running distinct tasks to get a feel for the notification system, then gradually scale up.
- Categorize your tasks: testing, refactoring, and documentation, which are relatively independent, are good candidates for parallel execution.
- Be mindful if one agent demands significant interactive time; pause other tasks to avoid getting bogged down in a global bottleneck.
Cherny's disclosure is less a definitive answer and more an experimental report from a practitioner on the front lines. It transforms the concept of 'one person doing a team's worth of work with AI' from a slogan into a tangible, actionable process. The next step will be to see if Anthropic integrates this parallel workflow into Claude Code's official features – after all, if the creator himself is using it this way, the tool team has every reason to pay close attention.
Comments
No comments yet
Be the first to comment