Stenobird

Episode

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

Podcast
Latent Space: The AI Engineer Podcast
Published
Mar 17, 2026
Duration seconds
5219
Processing state
processed
Canonical source
https://www.latent.space/p/felix-anthropic
Audio
https://api.substack.com/feed/podcast/191097767/72a8d1a77282f7fee0214a7442ccfd40.mp3
JSON
/v1/public/podcasts/latent-space-ai-engineer/episodes/why-anthropic-thinks-ai-should-have-its-own-computer-felix-rieseberg-of-claude-cowork-claude-code-desktop
Markdown
/podcast/latent-space-ai-engineer/why-anthropic-thinks-ai-should-have-its-own-computer-felix-rieseberg-of-claude-cowork-claude-code-desktop.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/latent-space-ai-engineer/episodes/why-anthropic-thinks-ai-should-have-its-own-computer-felix-rieseberg-of-claude-cowork-claude-code-desktop/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/latent-space-ai-engineer/why-anthropic-thinks-ai-should-have-its-own-computer-felix-rieseberg-of-claude-cowork-claude-code-desktop.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

The future of AI agents lies in moving from chat interfaces to trusted, local-first execution environments. Felix Rieseberg explores how Anthropic's Claude Cowork uses virtual machines to provide a safe, powerful sandbox for complex, non-coding tasks.

Topics

  • AI Agents
  • Claude Cowork
  • Anthropic
  • Virtual Machines
  • Agentic Workflows
  • Software Engineering
  • Automation
  • Local-first AI

Highlights

  • Main idea: The real frontier in AI is shifting from better chat interfaces to reliable, autonomous task execution
  • Practical takeaway: Using VM-based sandboxes allows agents to install tools and run scripts without constant human approval
  • Failure mode: Relying on 'approve every command' creates user exhaustion and breaks the flow of agentic workflows
  • Main idea: Execution is becoming cheap enough that teams can prototype by building multiple competing candidates rapidly
  • Practical takeaway: Local-first workflows are more powerful than cloud-only approaches because they can integrate with personal and professional context

Chapters

  1. 1:00 The Rise of Claude Cowork: An introduction to Claude Cowork and how it emerged from observing users applying Claude Code to non-coding knowledge work.
  2. 7:30 Cloud vs. Local Execution: A discussion on the trade-offs between running AI agents in the cloud versus on a user's local machine.
  3. 13:50 System Prompts and Agent Skills: How system prompts and specialized toolsets define the effectiveness of agentic workflows.
  4. 20:15 Evaluating Agent Capabilities: The process of testing and steering Claude through specific tools and prompt engineering.
  5. 33:20 AI for Knowledge Work: Moving beyond coding to use agents for data analysis, design-to-code, and automating manual digital tasks.
  6. 40:15 The Limits of AI Autonomy: A debate on where AI should and should not act autonomously, such as in sensitive financial transactions.
  7. 59:55 The Future of Privacy and Sandboxing: Exploring the use of virtual machines as a safety boundary and a way to handle sensitive personal data.