Anthropic product lead says Claude will shift to proactive features with safety-focused rollouts

Cat Wu says the next step for Claude is proactivity: Anthropic will pair anticipatory, agent — driven features with constrained, safety — first deployments and new supervision demands for teams.

At last week’s second annual Code with Claude conference in San Francisco, Cat Wu, Anthropic’s head of product for Claude Code and Cowork, outlined a push to make Claude more proactive while prioritizing safety in how features are rolled out. Her remarks signal that forthcoming updates will aim to anticipate user needs rather than just respond, a shift that matters for enterprises planning to embed Claude into workflows.

Wu, who joined Anthropic in August 2024, said she oversees new feature development across Claude variants, including the developer — focused Claude Code and the collaboration — oriented Cowork. She highlighted a frequent partnership with Boris Cherny, the creator of Claude Code-an internal duo observers have nicknamed Anthropic’s “Batman and Robin.” Anthropic released at least six models in the prior year and nearly as many so far this year; Wu added that future deployments may be structured differently from past releases.

The remarks come as Anthropic’s commercial momentum accelerates. The company is reportedly pursuing a funding round that could value it near $950 billion, compared with a reported $854 billion valuation for OpenAI in March. Anthropic cited a recent industry report saying business customers have increasingly preferred Claude over ChatGPT and that the company has quadrupled its business market share since May 2025. Wu described a product philosophy of “staying on the exponential,” meaning the team focuses on advancing model capabilities rather than reacting to competitors. She argued this approach shapes product road maps and deployment choices, pushing Anthropic to build forward — facing features that expand what Claude can do.

She also warned that wider adoption of agent — driven workflows will change how teams operate. Managers will need domain expertise to diagnose agent errors — for example, determining whether an agent misinterpreted instructions or whether a request was underspecified — and while routine, tedious tasks may be offloaded, organizations will require new skills for agent supervision and validation.

Wu pointed to Anthropic’s Glasswing initiative as an example of constrained, partner — focused deployments. Launched in April, Glasswing invited a small consortium, including Amazon, Apple, CrowdStrike and Microsoft, to test Mythos, Anthropic’s cybersecurity model. Unlike many other models, Mythos — designed to scan codebases for software vulnerabilities — was not given a general public release because the company said it feared the model could be weaponized. Throughout the conversation Wu reiterated that Anthropic wants to share advancing intelligence “as much as possible” while handling risks through careful deployment choices. She signaled that making Claude proactively anticipate user needs will require both safety guardrails and new product and management practices to ensure deployments are useful and secure.