Share with your CTO
Anthropic’s own engineering teams are running into a problem the company’s product helped create. Fiona Fung, the engineering leader who runs the Claude Code and Claude Cowork teams, says that as her engineers leaned harder into agentic coding workflows, they grew more isolated. Her fix is deliberately analog: programming lunches, hackathons, and shared maker-time blocks designed to let engineers watch each other work and exchange AI workflow patterns they’d otherwise never see. Claude Code has already become the dominant AI coding tool inside startups, according to a Business Insider survey of two dozen-plus founders and VCs.
What this means for your business
The engineers building the most advanced AI coding tool on the market couldn’t escape the isolation it produced. That’s not a soft cultural footnote. When your most productive developers are each running parallel agent swarms in their own private contexts, they stop cross-pollinating techniques. Institutional knowledge about how to get the most out of these tools doesn’t circulate. You end up with a team of capable individuals who have quietly diverged in their actual workflows.
Call it agent siloing: the state where each engineer’s most valuable productivity knowledge is locked inside their personal prompt history and agent configurations, invisible to teammates. It’s the opposite of what happened in the pair programming era, when technique spread through proximity. Fung’s observation that “everybody uses Claude Cowork so differently” is the tell. Divergent workflows in a team mean you’re not compounding on each other’s discoveries. You’re running dozens of isolated experiments with no shared lab notebook.
Anthropic’s response, structured in-person collaboration time, sounds like a retreat from the efficiency gains that justified the agentic tooling in the first place. It isn’t. It’s the recognition that the productivity ceiling for any individual engineer using AI agents is lower than the ceiling for a team that actively shares how it uses them. The signal worth watching: whether enterprises start treating AI workflow-sharing sessions as a formal engineering practice, not an informal nice-to-have.
Concept deep-dive: Agentic coding
Agentic coding means an AI system doesn’t just autocomplete lines of code. It autonomously plans, executes, and iterates across multi-step tasks, writing files, running tests, and correcting errors without the engineer touching each step. Think of it as the difference between a calculator and a contractor you brief once. The productivity gains are real, but the workflow becomes a private conversation between one engineer and their agent. That privacy is exactly why the knowledge-transfer problem Fung describes develops so quickly at scale.
Based on reporting from Anthropic Says Claude Code Is Making Its Engineers Lonely, originally published 2026-06-22 03:00:00.
