Cat Wu, head of product for Claude Code and Cowork at Anthropic, told TechCrunch that the next major shift in AI isn’t better models. It’s agents that act before they’re asked.
“I think the next big thing is proactivity,” Wu said in an interview at Anthropic’s second annual Code with Claude conference in San Francisco. “Claude understands what you work on, and just sets up some of these automations for you.”
The statement amounts to a product roadmap declaration. Anthropic’s Claude has moved through three phases: chatbot, coding tool, multi-agent workspace (Cowork). Wu is now describing a fourth: systems that monitor your work patterns, identify repetitive tasks, and create automations without a prompt.
The Three-Phase Trajectory
Wu joined Anthropic in August 2024. Since then, Claude has gone from answering questions to writing code to orchestrating multi-step workflows in Cowork. She described the progression in concrete terms: “Last year we were in this world of synchronous development. Right now, people are shifting to routines, so like automating responses to customer support tickets.”
The jump from routines to proactivity is architecturally significant. Routines still require a human to define the trigger and the action. Proactive agents observe, infer, and initiate. That’s a fundamentally different trust model.
Who Else Wants This
Anthropic isn’t alone in chasing proactivity. Google’s internal Remy project, reported by Droid Life on May 7, aims to turn the Gemini app into a 24/7 personal agent that can make purchases, send documents, and monitor events on a user’s behalf. OpenClaw has offered heartbeat-driven proactive agents since early 2026, where agents wake on a configurable schedule, check conditions, and act if thresholds are met, according to AIMultiple’s analysis of the platform.
The difference is architectural. Google’s Remy lives inside its first-party ecosystem: Gmail, Calendar, Search, Android. OpenClaw’s proactivity runs locally, using cron-like heartbeat polling against whatever tools the user has connected. Anthropic’s version would sit in the cloud, piggybacking on the workspace context that Cowork already captures.
Three platforms, three architectures, one bet: the next defensible moat in AI is not inference quality but ambient awareness.
The Management Problem
Wu addressed the obvious concern directly. If agents start anticipating and acting, does the human still need to understand the work?
“I think it is extremely hard to manage agents if you can’t do the job yourself,” she said. “Managing agents is actually very similar to being a manager of people, in the sense that you have to understand, like, why did the agent make this mistake? Did it misinterpret my instruction?”
This framing positions proactive agents as leverage, not replacement. But it sidesteps a harder question: proactive agents that silently create automations will produce errors that are harder to detect than ones generated by direct instruction. A bad response to a prompt is visible. A misconfigured automation running in the background for two weeks before anyone notices is a different failure mode entirely.
Why Now
The timing tracks with Anthropic’s broader momentum. Ramp data published the same day showed Anthropic surpassing OpenAI in business customer share for the first time: 34.4% of Ramp’s 50,000+ participating businesses now pay for Anthropic services, up from 9% in May 2025. OpenAI fell to 32.3%.
Ramp economist Ara Kharazian attributed the shift to Anthropic’s strategy: “Start with a very technical customer base, focus on their needs, really succeed in execution and then start broadening out through tools like Cowork.”
Proactivity is the next broadening move. Technical users already run Claude Code in complex workflows. Non-technical users are starting with Cowork. Proactive agents would collapse the gap between those two audiences by removing the need to specify what should be automated in the first place.
Wu also noted Anthropic’s model release pace: at least six in 2025, nearly as many in 2026 already. She referenced the Glasswing initiative, where Anthropic restricted access to its Mythos cybersecurity model to a small partner consortium, as an example of how deployment shapes might change as capabilities grow. Proactive agents with access to sensitive workflow data will face similar access control decisions.
The Competitive Framing That Wasn’t
One notable absence from the interview: Wu never mentioned competitors by name. “We don’t think about competitors,” she said. “If you do think about competitors, you end up being perpetually two weeks, or like, a month behind how fast you can execute.”
That’s a luxury afforded by Anthropic’s current position. When you’ve just overtaken OpenAI in business adoption and you’re raising at a potential $950 billion valuation (per Reuters), you can afford to talk about the frontier instead of the rearview mirror.
But the proactivity race has already started. Google’s Remy has a distribution advantage through Android and Workspace. OpenClaw has a head start with users who already run always-on agents locally. Anthropic’s pitch is that cloud-native context, captured through Cowork’s persistent workspace, gives it a richer signal for anticipation than either alternative.
The Trust Threshold
The real question isn’t whether proactive agents are technically possible. It’s whether users will grant them the ambient access required to observe, infer, and act without explicit instruction. Every proactive agent is, by definition, one that does something the user didn’t directly authorize at the moment of execution.
Wu’s framing of “everyone has this part of their life that’s really tedious” is the consumer pitch. The enterprise pitch is harder. CISOs who just spent six months building governance frameworks for prompted agents now face a product roadmap that wants to remove the prompt entirely.
Anthropic is betting that the value of anticipation will outweigh the friction of trust. The next six months will show whether that bet lands with the business customers who just made Claude their default.