Why Anthropic Thinks AI Should Have Its Own Computer — Felix Rieseberg of Claude Cowork & Claude Code Desktop

What It Covers

Felix Rieseberg, engineer at Anthropic, explains how Claude Cowork evolved from Claude Code into a VM-based knowledge work tool for non-terminal users. The conversation covers architecture decisions around local versus cloud compute, skills portability, sandbox security tradeoffs, and how Anthropic evaluates agent products against knowledge work tasks rather than coding benchmarks.

Key Questions Answered

• •VM-first architecture: Claude Cowork runs Claude Code inside a lightweight Linux virtual machine rather than directly on the host machine. This lets Claude install Python, Node.js, and any other dependency without requiring user approval or IT permission, while network ingress/egress controls remain strict. The VM collapses empty disk space so actual storage use is lower than macOS reports, though startup latency remains a real tradeoff versus running Claude Code natively.
• •Skills as portable markdown files: Cowork skills are plain markdown text files, not proprietary database entries. This makes them inherently portable — users can store them in any folder, sync via GitHub repos, or install plugin marketplaces by pointing to a GitHub URL. The unresolved gap is interpolating personal variables (preferred airports, phone numbers, folder paths) into otherwise shareable skill templates, which Rieseberg identifies as an unsolved industry problem worth building toward.
• •Eval against knowledge work, not coding benchmarks: Anthropic evaluates Cowork against knowledge work tasks — personal finance, mortgage management, legal office workflows — rather than standard coding suites used for Claude Code. The system prompt is tuned weekly based on these evals, steering Claude toward heavier use of the planning tool and ask-user-question tool for longer-horizon tasks where ambiguity is higher and wrong outputs are more costly to reverse.
• •Scaffolding has diminishing returns as models improve: Rieseberg argues that investing heavily in task-specific scaffolding carries risk because each model generation reduces the gap between raw capability and scaffolded output. His current preference is to maximize tool access and safety boundaries rather than build elaborate correction layers, then let model improvements handle the rest. He observes this pattern already in MCP servers being partially superseded by skills as models generalize better.
• •Local compute retains strategic value: Rieseberg pushes back on cloud-first assumptions by noting that moving all agent work to the cloud requires solving authentication, data privacy, and permission synchronization simultaneously. A concrete example: reading Chrome cookies to enable cloud-based browser sessions would trigger bank account lockouts from location anomalies. Keeping Claude on the local machine sidesteps these problems while giving it access to the same file system, applications, and credentials the user already has.
• •Build all candidates, then pick: Anthropic's current product development process skips traditional spec-then-build cycles in favor of rapidly constructing multiple prototypes simultaneously and selecting the best performer with a small focus group. Rieseberg credits this shift — enabled by cheap execution — with Cowork's ten-day build timeline, while clarifying that existing primitives like Claude Code, the VM framework, and internal prototypes accumulated over eighteen months provided the actual foundation.