SAN FRANCISCO - Anthropic doesn't have a long-term road map for Claude Code, its agentic software development tool, but the company is betting that such a plan would be rendered moot by rapidly improving model capabilities and whatever developers decide they want next week. That's the takeaway from a 30-minute conversation Ars had with Cat Wu, Anthropic's head of product for Claude Code.
Last week, in a three-level car rental parking garage meticulously converted into an event space in downtown San Francisco, Anthropic held its second annual Code with Claude developer conference. The single-day event included a keynote introducing new features for Managed Agents and announcing a compute deal with SpaceX. That compute deal came with a doubling of usage limits for Claude Code users on Pro and Max plans - a response to significant user frustration about a compute crunch, especially in recent weeks.
Anthropic's products - especially Claude Code - have seen runaway popularity. "We tried to plan very well for a world of 10x growth per year," CEO Dario Amodei said on stage. "And yet we saw 80x, and so that is the reason we have had difficulties with compute." User growth was accompanied by a shift away from simple chat interfaces to complex, multi-agent workflows that are many times more demanding. During the crunch, Anthropic tested solutions like enforcing stricter limits during peak hours or removing Claude Code from its cheaper subscription plan.
Over the past year, Anthropic has released a plethora of new features, products, and surfaces for interacting with its models. Claude Code went from CLI to IDE to desktop, and new tools for managing multiple agents were rolled out. The pace has been intense and chaotic at times. Meanwhile, competitors like OpenAI's Codex, GitHub Copilot, Cursor IDE, and Augment Code are rolling out their own products, sometimes with hooks like more explicit context they claim leads to better results.
As head of product for Claude Code, Wu works closely with its creator, Boris Cherny, to identify which features to prioritize. She does not oversee the models, but the product strategy she describes makes a big bet that models will continue to improve so rapidly that planning what a product like Claude Code should look like in the future is essentially futile. The Claude Code team goes through development cycles of just a week or so, rolling out features in a Wild West of experimentation.
When asked whether the CLI remains the center of gravity, Wu noted that usage is split across all surfaces. "The center of gravity is still the CLI," she said. "It's still the one that has the most power-user features, it's where most of our features land first, and it's also just the fastest for us to iterate on." However, she observed a gradual shift toward desktop as developers went from managing one agent to six terminal tabs and eventually decided reading ten tabs was not their idea of a good time.
On the question of whether Anthropic might consolidate its many surfaces, Wu described a progression: most start in CLI or IDE, then graduate to desktop for multi-agent management, then want routines that watch Slack channels automatically. "All the products are just a way to help you more easily elicit the intelligence of the models," she said. "We actually remove scaffolds. We remove parts of the system prompt and tool descriptions over time as models get smarter." She can envision a world where everything collapses back to a text box if the model is "always right," but for now, they need all the tools.
Wu cited Richard Sutton's 2019 essay "The Bitter Lesson" as a guiding principle, which argues that general-purpose methods that scale with compute ultimately win out over domain-specific structures. "We're pretty humble about not knowing exactly what the right form factor is but encouraging our teams to explore that as much as possible," she said.
On compute limits, Wu noted that while plugins exist for semantic codebase navigation, they haven't found a measurable improvement in performance from adding structured data. "We generally lean more toward shipping a leaner harness with fewer opinionated tools and just letting developers add their own if they want," she said. Token efficiency is always on their minds, but "the most important thing is just maintaining intelligence" - that's the north star, not token efficiency.
As for what's next, Wu sees Claude potentially anticipating what developers want: monitoring GitHub issues, Slack feedback, and Twitter for bugs or feature requests on whatever feature someone is working on, then automatically setting up routines and proposing actions. "It's actually not that far away," she said, suggesting engineers might soon not need to set up automations themselves - Claude will just decide to listen for feedback and propose what to do today.