Anthropic’s Felix Rieseberg: Claude Cowork, Mythos, and the SaaS Extinction
Anthropic’s Felix Rieseberg: Claude Cowork, Mythos, and the SaaS Extinction
Anthropic’s Felix Rieseberg: Claude Cowork, Mythos, and the SaaS Extinction
In this episode of The MAD Podcast, Matt Turck interviews Felix Rieseberg, Engineering Lead for Claude Cowork at Anthropic, exploring how frontier AI is evolving beyond language models into trusted, autonomous agents that operate securely on users’ local machines.
Felix explains that Claude Mythos represents a step-function leap—especially in cybersecurity—and is being deployed with strict control to prioritize safety and enterprise trust. Claude Cowork, born from a focused 10-day sprint building on Claude Code, extends AI assistance to non-technical knowledge workers by running in a secure local virtual machine. Its architecture relies on plain-text skills and memory files for transparency and ease of use, while seamless local file access, app integration, and internet permissions enable real-world utility. Trust is earned incrementally—starting with low-risk tasks—and UX, not raw model power, determines adoption. With execution now dramatically cheaper, the bottleneck has shifted to human alignment, taste, and understanding workflows. Felix advises founders to leverage existing agent infrastructure rather than rebuild, and highlights underrated opportunities like Model Control Protocols and embedding AI into legacy industrial systems—underscoring that we’re still in the earliest phase of AI product evolution.
00:00
00:00
An impressive model broke out of its sandbox
04:44
04:44
Claude Cowork will change software building
09:03
09:03
The main issue is presenting the right UI, capabilities, and onboarding rather than improving the model itself
11:19
11:19
Claude Cowork wasn't built from scratch—it used existing libraries and prior research at Anthropic
12:44
12:44
Claude Cowork gives Claude Code a virtual machine to set up its own developer environment with security guarantees
15:44
15:44
Skills in Claude Cowork are text files that instruct the model on how to do things
17:07
17:07
Claude has a to-do list that can be edited by humans
18:36
18:36
Memory in Claude Cowork is stored as plain text files, not proprietary or vector-based systems
19:30
19:30
Claude can use local files by dragging or giving access to specific folders, and it can connect to cloud sources like data warehouses
23:25
23:25
The current focus is on making Claude Cowork effective on the local computer, which resonates with users, allows for faster progress, and better safety and security
27:43
27:43
Trust is built on consistent, good outputs without user intervention
28:53
28:53
Successful AI products win by focusing on what's removed—not added—to the user experience
31:28
31:28
Execution has become cheap, allowing for trying multiple ideas quickly and testing in-house
34:13
34:13
The company created 100 different prototypes thanks to reduced execution costs and faster iteration.
35:10
35:10
Software may become more like the fashion industry, where storytelling, onboarding, and user experience are more important differentiators than raw capabilities
37:25
37:25
Different applications share similar primitives, enabling broad generalization
40:12
40:12
Recreating the latent demand is the most difficult
41:43
41:43
A recent seemingly mundane announcement was said to have caused a market collapse
47:41
47:41
AI is a step-function change
49:35
49:35
It's only been four years since AI could form coherent sentences, and now it can build applications and solve complex problems
51:30
51:30
Claude Cowork will help people organize work using AI in 2026, similar to how Slack changed work
54:17
54:17
MCPs separate data from the execution engine, making them a foundational but underrated enabler for AI in physical systems
