Anthropic's Co-Work: Virtual Machines, Skills, and Local-First AI

The Unseen Architecture: How Anthropic's Co-Work Reimagines AI's Desktop Presence

The conversation with Felix Rieseberg, Member of Technical Staff at Anthropic, reveals a profound shift in how we should conceive of AI products. Beyond the immediate allure of enhanced chat capabilities, the true frontier lies in trusted task execution, a domain where AI agents are granted their own computational space. This episode unveils the non-obvious consequence of this approach: a re-evaluation of the local computer's role, a reimagining of user-AI collaboration through sandboxed environments, and the creation of "skills" that promise portable, personalized automation. Those who grasp these downstream implications will gain a significant advantage in building and leveraging AI tools that move beyond mere conversation to genuine, independent action. This insight is crucial for product builders, AI researchers, and anyone seeking to understand the next wave of human-computer interaction.

The Virtual Machine as a Gateway to Autonomy

The core of Anthropic's Co-Work, as explained by Felix Rieseberg, isn't just a more user-friendly interface for Claude Code; it's a fundamental architectural choice: running Claude within a dedicated virtual machine (VM). This isn't merely a technical detail; it's the linchpin for enabling AI autonomy and safety. By providing Claude with its "own computer," the system unlocks capabilities previously constrained by the host operating system's limitations and security protocols. This allows Claude to install necessary tools, run scripts, and operate with a degree of independence that a traditional cloud-based chat interface simply cannot achieve.

The immediate benefit is obvious: a more capable AI. However, the deeper consequence is the shift in the user's relationship with the AI. Instead of a constant back-and-forth of prompts and approvals, the VM acts as a controlled environment where Claude can operate more autonomously. This isn't about relinquishing control entirely, but about creating a more sophisticated middle ground between "uselessly safe" and "dangerously autonomous." The VM serves as a sandbox, a dedicated space where Claude can experiment, learn, and execute tasks without the constant risk of impacting the user's primary system. This architectural decision directly addresses the "approval fatigue" that plagues many current AI interactions, where users are bogged down by constant prompts to confirm every minor action.

"It is quite powerful to give Claude its own computer that is like generally a good idea. And in terms of architecture and UX and everything else that we've been working on in Anthropic, it often is quite useful for you to like anthropomorphize, um, Claude aggressively and just be like, 'This is a person.'"

-- Felix Rieseberg

This anthropomorphization, facilitated by the VM, allows users to conceptualize Claude as a more independent agent, capable of managing its own environment and tools. The downstream effect is a more natural delegation of complex, multi-step tasks. Conventional wisdom might suggest that the path to AI utility is solely through more powerful models or more intricate prompt engineering. Rieseberg's perspective, however, highlights how the environment in which the AI operates is equally critical. By providing a dedicated VM, Anthropic is not just offering a better chat window; they are building a platform for AI agents that can truly do things, rather than just say things. This distinction is crucial for competitive advantage, as it allows for the development of AI workflows that are not only more powerful but also more trustworthy due to the inherent safety mechanisms of the sandboxed environment.

Skills: The Portable Primitives of Personalized Automation

The concept of "skills" in Co-Work represents another significant departure from traditional AI product design. Instead of relying on rigid, predefined tool schemas or complex API integrations, skills are presented as lightweight, markdown-based instructions. This approach democratizes the creation of personalized automation, allowing users to define reusable workflows that can be easily shared and adapted. The immediate payoff is the ability to automate repetitive tasks, from organizing files to uploading videos.

However, the second-order effect of skills is far more profound: they become the portable primitives for personalized AI. Rieseberg emphasizes that skills are designed to be file-based and text-native, making them inherently portable across different agentic products. This is a critical insight that challenges the notion of proprietary AI ecosystems. If a skill can be easily transferred from one AI harness to another, it shifts the value proposition from vendor lock-in to interoperability. The "hardest questions ahead," as Rieseberg notes, are about autonomy, safety, and portability. Skills directly address portability by decoupling the AI's capabilities from the specific platform it's running on.

"My personal belief is that the hackability and the extensibility has like played a pretty big role, right? You can hook in Visual Studio Code that like almost any workload, it's so easy to hack on, so easy to put extensions for it. And I think Co-Work might be hitting a similar thing."

-- Felix Rieseberg

The real competitive advantage here lies in the delayed payoff of a truly interoperable skill ecosystem. While immediate automation is valuable, the long-term implication is a user-defined layer of intelligence that can move with them across different AI tools. This requires a fundamental shift in thinking: instead of building monolithic AI applications, the focus moves towards creating robust, general-purpose primitives like skills that can be combined and recombined in novel ways. Conventional wisdom often dictates building highly specialized AI solutions for specific verticals. Rieseberg suggests that a more durable strategy is to build powerful, general primitives that can be adapted to any vertical, absorbing narrow use cases as models improve and these primitives become more sophisticated. This approach, while requiring patience and a focus on foundational elements, creates a moat that is difficult for competitors to replicate.

The Local-First Frontier: Reclaiming the Desktop

A recurring theme in the conversation is Anthropic's bet on "local-first" agent workflows, challenging the prevailing Silicon Valley narrative that everything must migrate to the cloud. Rieseberg argues that the local computer, often overlooked, remains a critical frontier for AI products. The argument is compelling: why should an AI agent that is meant to be intimately helpful be confined to a remote server, requiring constant permissions and potentially creating privacy concerns? By bringing Claude "where you work" -- on your local machine, within a VM -- Co-Work offers a unique blend of power and intimacy.

The immediate benefit is an AI that can directly interact with the user's existing tools and data. The example of Co-Work integrating with Chrome allows it to see and manipulate the browser's DOM, enabling complex workflows like programmatic uploading to YouTube. This direct access, however, raises questions about security and trust. The VM architecture is the crucial mitigating factor, providing a controlled environment that balances capability with safety. The downstream effect is a more seamless integration of AI into daily workflows, reducing the friction associated with cloud-based solutions that often require extensive setup and permission granting.

"I generally believe that Silicon Valley overall is undervaluing the local computer. And my default argument for that is always how come we're all using MacBooks and not like an iPad or a Chromebook?"

-- Felix Rieseberg

The delayed payoff of this local-first approach is the creation of AI agents that are deeply integrated with a user's personal data and workflows, fostering a level of trust that remote, cloud-based solutions may struggle to achieve. This requires a long-term investment in building secure, performant local execution environments. Conventional wisdom might push towards cloud-native solutions for scalability and ease of access, but Rieseberg's argument suggests that for certain types of AI tasks, particularly those requiring deep integration with personal data and existing applications, the local computer offers a more powerful and potentially more private alternative. The competitive advantage lies in building AI that feels like an extension of the user's own digital environment, rather than a separate, external service. This requires a willingness to invest in the complexities of local execution, a path many in the industry have shied away from.

Key Action Items

• Embrace the VM as a Core Architectural Component: For AI products aiming for task execution and autonomy, prioritize building within a sandboxed virtual machine environment to enhance safety and capability. (Immediate Action)
• Develop and Promote Portable Skills: Invest in creating a flexible, text-native "skills" system that allows users to define and transfer personalized automation workflows across different AI agents. (Immediate Action, Pays off in 6-12 months)
• Prioritize Local-First Integration: Design AI agents that can deeply integrate with the user's local machine and existing applications, leveraging the power of the desktop rather than solely relying on cloud infrastructure. (Long-term Investment, Pays off in 12-18 months)
• Rethink User-AI Interaction: Move beyond constant prompt-and-approval loops by designing systems that allow for more autonomous AI action within safe, controlled environments. (Strategic Shift, Pays off in 12-24 months)
• Focus on Generalizable Primitives: Instead of hyper-specialized AI verticals, build robust, general-purpose AI capabilities (like skills) that can adapt to a wide range of tasks and industries. (Strategic Shift, Pays off in 18-36 months)
• Invest in Robust Local Execution: Dedicate resources to optimizing the performance and reliability of AI running directly on user machines, acknowledging the unique challenges and opportunities of local-first development. (Long-term Investment, Pays off in 12-18 months)
• Champion Interoperability: Design AI tools and workflows that are inherently interoperable, allowing users to leverage their own custom automations and data across different platforms. (Strategic Shift, Pays off in 18-24 months)