Anthropic has no long-term road map for Claude Code, the company's product lead for the agentic coding tool said this month, and that is deliberate. Cat Wu, who runs product for Claude Code alongside its creator Boris Cherny, told Ars Technica in a 30-minute interview at the second annual Code with Claude developer conference in San Francisco that the team ships new surfaces on week-long cycles and bets the underlying model will outrun any plan it could write down. The conference also produced a doubling of usage limits on the Pro and Max plans and a compute deal with SpaceX, both aimed at a capacity crunch that has bruised paying users in recent weeks.
The scale problem is now on the record. "We tried to plan very well for a world of 10x growth per year," Anthropic CEO Dario Amodei said at the event. "And yet we saw 80x, and so that is the reason we have had difficulties with compute." The 80x figure refers to demand for Claude Code and related agentic workloads, which consume far more tokens per session than a chat turn.
To survive the crunch, Anthropic has tested stricter limits during peak hours and pulled Claude Code from its cheaper subscription tiers entirely at points. The doubled limits on Pro and Max, paired with the SpaceX compute, are the company's answer for now. The event itself was held in a three-level car rental parking garage in downtown San Francisco that Anthropic converted into a venue.
Key facts
- 01Anthropic planned for 10x annual growth on Claude Code but saw 80x, CEO Dario Amodei said on stage at Code with Claude 2026.
- 02The company doubled usage limits for Claude Code on Pro and Max plans alongside a new compute deal with SpaceX.
- 03Product lead Cat Wu says the team ships on roughly week-long cycles, going from conviction to product shift in about seven days.
- 04Internal usage has shifted from one agent per developer to roughly six terminal tabs, pushing more users toward the desktop app.
- 05Anthropic is removing scaffolds, system prompt text, and tool descriptions from Claude Code as underlying models get smarter.
Wu's product philosophy is to chase signal, not the calendar. "We have no grand plan," she told Ars, laughing. "We try to go from conviction to a product shift pretty quickly, ideally in a week or so." That has produced a sprawling surface area: a CLI, an IDE integration, a desktop app, and higher-level products including Cowork and Managed Agents.
“We tried to plan very well for a world of 10x growth per year. And yet we saw 80x, and so that is the reason we have had difficulties with compute.”— Jaeden Schafer
Asked where the center of gravity sits, Wu was specific. "We actually find that every developer has a different preference, so our usage is pretty split between all these," she said. "The center of gravity is still the CLI. 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."
Internal usage tells the story of how agentic coding has scaled inside Anthropic itself. A year ago, Wu said, most engineers had one agent running. Over the course of the year, that grew to roughly six terminal tabs per developer, and ten tabs is now the threshold where people give up and switch to the desktop app for a richer view. "I understand why people have graphical interfaces," Wu said.
The strategic bet underneath all of this is that scaffolding around the model is temporary. "We actually remove scaffolds. We remove parts of the system prompt and tool descriptions over time as models get smarter," Wu said. She cited Richard Sutton's 2019 essay The Bitter Lesson — which argues that general methods scaling with compute beat hand-engineered structure — as a guiding principle for the team.
Taken to its conclusion, the form factor itself becomes negotiable. "I can definitely see a world where maybe we all just collapse back to the text box again because maybe the model is just always right," Wu said. The near-term feature she sees coming is proactive monitoring: Claude deciding on its own to watch a GitHub issue list, a Slack channel, or a Twitter feed for feedback on a feature an engineer is shipping, then surfacing what matters without an automation being set up by hand.
Anthropic is operating in a crowded field. OpenAI's Codex now ships inside the ChatGPT mobile app, GitHub Copilot has been pushed harder by Microsoft, Cursor's IDE has a large paying base, and Augment Code is selling explicit-context retrieval as a differentiator. Microsoft recently canceled internal Claude Code licenses and moved engineers to its own Copilot CLI, a sign that the agentic coding market is now competitive enough for large customers to pick sides.
The risks in Wu's approach are real. Shipping new surfaces on week-long cycles produces sprawl: a CLI, an IDE plugin, a desktop client, Cowork, and Managed Agents all need to be maintained and explained to users who may not know which one to open. The compute shortfall is also not solved — doubled limits and one new supplier do not close an 80x-versus-10x gap on their own, and the next quarter of growth will test whether the SpaceX deal and any successors keep pace.
The thesis Wu is articulating is the cleanest statement yet of how Anthropic intends to compete on coding: keep the harness lean, trust the model curve, and let users vote with their workflows on which surface survives. If the next Claude generation closes the reliability gap on long-horizon agentic tasks, the desktop app, the IDE, and Cowork all become optional skins on the same text box — and the moat is whoever has the most agentic compute to sell. If it does not, Anthropic will be running a product portfolio it explicitly declined to plan for, against rivals who did.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
