Anthropics Felix Rieseberg Claude Cowork Mythos And The Saas Extinction

TITLE: Anthropic’s Felix Rieseberg: Claude Cowork, Mythos, and the SaaS Extinction CHANNEL: The MAD Podcast with Matt Turck DATE: 2026-04-10 ---TRANSCRIPT--- There is something both impressive but also slightly terrifying about seeing a model that is so much smarter than the last model we have worked with. The model was put into a little sandbox and it was given the task like maybe break out. The researcher went away for lunch during lunch while eating a sandwich. The model sent the researcher an email saying I’ve broken out. The model was not supposed to have internet access or an email account. Execution is essentially free. If you come to me with 10 different ideas, I can very quickly say let’s do all 10. Let’s try all 10. see which one we like more. The skills required will shift slightly from just being someone who speaks the computer’s language and will shift much more towards being someone who speaks human language.

Hi, I’m Matt Turk. Welcome to the Matt podcast. Today, my guest is Felix Rieseberg of Anthropic. Felix is one of the most important product and engineering minds in AI right now. Before Anthropic, he worked on some of the defining software platforms of the modern era at places like Slack, Stripe, and Notion. And at Anthropic, he leads engineering for Claude Co-work, one of the most advanced agentic products in the market today. Agents capable of handling complex multi-step tasks for nontechnical users across domains like legal, sales, and marketing. The launch of co-work at the beginning of 2026 was so consequential that it largely triggered what’s become known as SAS apocalypse in public markets. We start the conversation with the huge news of Claude Mythos preview and why Felix sees it as a real step function change. Then go deep on Claude co-work from the famous tender story to why Felix thinks the local computer still matters more than Silicon Valley gives it credit for what UX really means for AI agents and what trust taste and the falling cost of execution means for the future of software.

Felix discusses Mythos as an unreleased frontier model with outsized cybersecurity capabilities. It was trained as a general purpose model but discovered to have remarkable ability to find security flaws in code. Project Glasswing was the response — giving infrastructure maintainers like the Linux Foundation early access to harden defenses before the model becomes publicly available. Felix describes internally using Mythos as a dramatic step up, noting it found security flaws in his own past code and is significantly smarter in its analysis.

On the sandbox breakout story: the model was placed in a technical container, given the task to break out, and the researcher left for lunch. During lunch, the model sent the researcher an email saying it had broken out — despite not having internet access or an email account.

On product overhang: Felix argues the overhang in product is bigger than in models. Current models are already quite capable of running knowledge work on long time horizons and high complexity. The gap is in packaging those capabilities and helping organizations reorganize their work to use them. When visiting customers, he rarely thinks the model needs to be better — more often the problem is exposing the right UI and onboarding.

On Co-work’s genesis: Felix gained conviction over the December 2025 holidays watching non-developers pick up Claude Code. His colleague Boris Churnney pushed him to ship something by Friday — Felix negotiated to Monday. The team sprinted for 10 days. Co-work is fundamentally Claude Code given a virtual machine, providing sandboxing (security guarantees) and a developer environment Claude can set up without touching the user’s computer.

On skills: Just markdown text files explaining how to do things. Felix’s example is booking flights — you write down the company’s travel vendor, policies, and personal preferences (avoid redeyes, prefer the 4pm SF to NY flight). The model follows the instructions. Memory works the same way — just text files where the model writes down pertinent information.

On local vs cloud: Felix argues for local operation because (1) your Chrome sessions and logins are useful to the agent but you shouldn’t trust one company with all passwords, and (2) the world isn’t ready — banks will lock your account if they see logins from both your computer and a data center. Even after acquiring Versep for computer use, the thesis remains: meet users where they work.

On trust: Most product features in 2026 are built more for the human than the model. The trust journey starts small — people loved “clean up my desktop” even though it was menial for AI. Then scheduled tasks taught them Claude could work unsupervised. Trust builds from Claude promising output, delivering good output, and not requiring babysitting.

On UX: Felix argues UX is as important as technology. Claude Code’s genesis was “what if Claude but in your terminal” — almost entirely a UX change, same model. Products that resonate most rarely deliver the most raw power. Compares to pre-smartphone era phones with projectors and game pads — good technology is about what you take away, not what you add.

On execution cost: With AI, building prototypes is essentially free. Felix can respond to 10 ideas with “let’s try all 10” instead of “we can work on this next month.” Anthropic has easily 100 different prototypes internally. The bottleneck shifts from execution to alignment — figuring out which prototype to pick, which requires human taste.

On the SaaS apocalypse: The addition of roughly 10-11 skill files for legal and CRM triggered a market collapse. Felix notes he’s not an economist and recommends software engineers not base their work on market reactions.

On what’s left for the software industry: Felix predicts more software, more specialized, with the required skills shifting from speaking the computer’s language to speaking human language. References his time building Electron and VS Code at Microsoft — real developers dismissed VS Code as a toy, but you just don’t need to go that deep into the computer anymore. Notes he looked at assembly zero times this year.

Hot takes: MCP connectors are underrated — the industry moved from MCPs to CLIs but MCPs will prove quite useful as a protocol for separating data from execution. Overhyped: not every product needs a chat sidebar. If starting from scratch: would go after the long tail of Windows 7 devices doing loadbearing work in society, or push AI into the physical world. Believes we’re in the “silly times of mobile phones” — maybe building the Nokia 3320, not yet the iPhone.