Cat Wu, head of product for Claude Code, details how persistent memory and compressed development cycles are changing what AI coding agents can actually do.
Cat Wu, Anthropic’s Head of Product for Claude Code, recently laid out the company’s approach to building what amounts to a context-rich harness for AI coding agents. The core idea: give AI tools persistent memory, shared workflows, and enough situational awareness to handle complex development tasks.
For crypto developers writing security-sensitive smart contracts or auditing DeFi protocols, that shift from “fancy autocomplete” to “contextually aware coding partner” matters more than most people realize.
From prompts to persistent context
Wu joined Anthropic in August 2024 to lead product development for Claude Code and a related product called Cowork. Since then, the team has shipped features like Claude Tag, which introduces persistent context and memory layers for shared workspace management.
A July 2026 discussion featuring Wu alongside engineer Thariq Shihipar dove into Claude Fable and Claude Tag, two features that push this persistent-context approach further. The emphasis wasn’t on making the underlying model smarter, necessarily. It was on making the harness, the orchestration layer sitting between the developer and the model, dramatically more capable.
Speed as a product philosophy
One of the more striking details from Wu’s comments: Anthropic’s product development timelines have compressed from the traditional six-to-twelve month planning cycles down to as short as one week. The compression comes from what Wu describes as AI-native workflows, where the tools the team is building are also the tools the team uses to build.
Where crypto meets AI coding agents
Here’s the thing. None of Anthropic’s recent product announcements mention crypto tokens, blockchain projects, or DeFi partnerships. Wu’s discussions have stayed firmly in the lane of general-purpose developer tooling.
Context-aware AI coding agents that maintain persistent memory across sessions could meaningfully improve protocol audits and security reviews. Instead of an AI tool that forgets the codebase structure between prompts, imagine one that understands the full architecture of a DeFi protocol, remembers previous audit findings, and can flag patterns that resemble known vulnerability classes. That’s roughly what Claude Tag’s persistent context layer enables, applied to a specific domain.
