Most people struggle to keep one conversation going at a time. Boris Cherny, the head of Claude Code at Anthropic, runs five Claude instances simultaneously in his terminal, with another five to ten sessions open on the claude.ai platform. His approach to AI-assisted coding looks less like a traditional developer workflow and more like an air traffic controller managing a dozen planes at once.

From prototype to productivity engine

Claude Code started life as a terminal-based prototype in late 2024. It was never designed to be a polished product with guardrails and dropdown menus. Instead, it was built to let developers interact with Claude directly from the command line, moving tasks fluidly between local instances and web-based sessions.

Cherny’s January 2026 thread detailing his personal workflow went massively viral, amassing millions of views across social media. The thread walked through how he orchestrates multiple Claude sessions, assigns different tasks to different instances, and shifts work between local and cloud contexts depending on what each job requires.

The results inside Anthropic have been striking. Productivity per engineer reportedly increased by nearly 70% following widespread adoption of Claude Code. That number becomes even more impressive when you consider that Anthropic’s workforce tripled in size over the same period.

The shift from prompts to loops

In discussions on Lenny’s Podcast and at the Sequoia AI Ascent event, Cherny described a progression that most developers haven’t caught up with yet.

The first phase was direct interaction. You talk to the AI, it responds, you iterate.

The second phase involves what Cherny calls autonomous loops. Rather than manually guiding each step, developers set up tasks that Claude can work through independently, checking back when the loop completes or hits a decision point that requires human judgment.

The third phase is where things get genuinely wild. Cherny described running “a few thousand” overnight sub-agents for complex tasks.

According to Cherny, 80-90% of Claude Code itself is now generated through these advanced methods. A substantial portion of Anthropic’s broader engineering output follows the same pattern.

What this means for investors

The productivity numbers coming out of Anthropic’s Claude Code adoption deserve serious attention from anyone watching the AI tools market. A nearly 70% increase in output per engineer, sustained through a period of rapid team expansion, is the kind of metric that reshapes how companies think about R&D spending.

The shift toward autonomous agent orchestration also opens a different kind of investment lens. Companies building infrastructure for managing thousands of concurrent AI agents, monitoring their output, and integrating their work into existing codebases are addressing a bottleneck that Cherny’s workflow makes obvious. Running a few thousand overnight sub-agents sounds impressive until you realize someone has to review, test, and deploy what they produce.