A Fireside Chat with Cat and Thariq from the Claude Code team

This piece is a summary of a fireside chat with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team, hosted by Simon Willison at the AI Engineer World's Fair. The discussion covers how coding agents like Claude Code, Fable, and Claude Tag are changing daily software development, team collaboration, and internal workflows at Anthropic.

A central theme is how coding agents have shifted engineering work from manual execution to delegation and product direction. Because building features is much faster, product taste and business sense have become more valuable than traditional long-form specs. The team uses an internal tool called Claude Tag, a multiplayer and proactive Slack integration that now handles 65 percent of product engineering pull requests. Anthropic relies heavily on a culture of internal dogfooding, known as ant fooding, where features are tested internally and must meet strict user retention bars before public release.

The conversation also highlights changes in prompt engineering and code review. Modern models perform better with significantly leaner system prompts, with the Claude Code system prompt shrinking by 80 percent after removing bloated examples and negative constraints. While critical core changes still require manual human review, the team increasingly trusts automated code review for outer layers of the product, backed by extensive internal and external evaluations to catch regressions and security risks.