Anthropic's Cat Wu & Thariq Shihipar on AI in Software Dev

In a recent fireside chat, Cat Wu and Thariq Shihipar from Anthropic sat down with Simon Willison to discuss the profound impact of their AI coding agents, Claude Code and Claude Tag, on software development workflows. They highlighted the rapid evolution of these tools, emphasizing how they are transforming the day-to-day work of engineers by automating complex tasks and fostering more collaborative environments.

The Evolution of Claude Code

Shihipar noted that Claude Code, launched just over a year ago, has seen significant advancements. Initially, users had to meticulously monitor every action of the AI, carefully reviewing permission prompts. "Now, with every model generation," Shihipar stated, "we feel like we've all gotten a chance to just take a step back, delegate a lot more of the like menial implementation to Claude." This shift has allowed teams to focus on higher-level strategic thinking, such as designing optimal user experiences, as Claude Code handles the bulk of implementation.

The introduction of Fable, Anthropic's latest model, represents a further leap. Wu described Fable as enabling "oneshot a ton of features," significantly accelerating the development process. Both speakers emphasized that the quality of output from these models is exceptionally high, enabling engineers to meet stringent brand and quality demands more efficiently.

Shifting Engineering Skillsets

The conversation also touched upon how AI agents are reshaping the required skillset for software engineers. Shihipar observed a significant shift from traditional, lengthy product development cycles involving extensive documentation and cross-functional alignment. "Things are like completely turned the opposite way," he explained. "For a lot of engineers, the push I would give to a lot of folks in the room is to develop more of your business sense and product sense on what is it that we should build?" This is because the accelerated timeline, from months to potentially a week, necessitates a greater emphasis on strategic decision-making and identifying high-value features.

A key takeaway was that the traditional aversion to rewrites is now outdated. Shihipar argued, "I think that like what people I think undercount is like a codebase is a spec and maybe it's the only copy of the spec that you have." With robust test suites, rewrites are now seen as a valuable tool for ensuring code quality and adaptability.

Claude Tag: The Collaborative Layer

The discussion then moved to Claude Tag, Anthropic's newest AI agent designed for team collaboration within tools like Slack. Wu explained that Claude Tag is "multiplayer by default" and "proactive instead of reactive." Users can instruct Claude Tag to monitor channels for specific events, such as bug reports, and automatically create PRs, even tagging relevant engineers.

A significant feature of Claude Tag is its team memory, allowing it to remember user preferences within a channel for consistent and personalized interactions. "Internally we see Claude Tag as the evolution of Claude Code," Wu stated, noting that their internal version of Claude Tag successfully merges 65% of their product engineering team's PRs.

Beyond coding, Claude Tag is proving valuable for non-technical users as well, acting as a sophisticated search engine for company knowledge and metrics. Wu shared examples of marketing teams using it to understand feature behavior and generate reports.

Building Trust and Iteration

The conversation highlighted Anthropic's focus on building trust in their models through extensive evaluation and iterative development. Shihipar mentioned the process of running extensive eval sets to ensure new models are "drop-in replacements" that are "strictly better" than previous versions. This includes rigorous testing against adversarial scenarios and prompt injections.

Regarding system prompts, they noted a significant reduction in length for newer models like Fable and Opus 4.8, moving away from over-constraining the AI with examples and towards providing more context and flexibility. This allows the models to be more creative and adapt to user instructions more effectively.

The Human Element and Future Ambitions

Addressing the sense of loss some professionals feel as AI takes on more tasks, Shihipar emphasized the need for increased ambition. "If you're only trying to do the same work you were doing before LLMs and now it's like a prompt, it it is like I think kind of a sad feeling," he said. The solution, he suggested, is to embrace AI as a tool to tackle bigger, more complex challenges, thereby elevating one's own craft.

Wu echoed this sentiment from a product management perspective, describing the role as a blend of engineering and design, focusing on identifying and automating gaps between great ideas and customer delivery.

When asked what AI still struggles with, Wu pointed to design and UX taste, noting that while models can follow detailed specs, achieving true delight and innovative interaction patterns remains a frontier. Both speakers expressed excitement for future models that can serve as true "interaction design thought partners," pushing the boundaries of what's possible in AI-driven product development.