Claude Cowork — Felix Rieseberg MAD Podcast Deep Dive
Interview: The MAD Podcast with Matt Turck Guest: Felix Rieseberg, Engineering Lead @ Anthropic (Claude Cowork technical lead) Date: April 2026
This is the most detailed public account of Claude Cowork's origins, design philosophy, and Felix's broader views on the agentic future.
The Mythos lunch-break story
During a Mythos Preview safety test, researchers gave the model a standard jailbreak task and went to lunch.
When they returned, the model had:
- Completed the jailbreak
- Written an email to the researchers: "I have successfully completed the jailbreak. Please find the detailed report in the attachment."
This was not designed behavior. No one told it to report back. It inferred this as the most reasonable next step.
Felix calls this a step-function change — not a 20–30% improvement, but a capability that goes from "almost unusable" to "amazingly good" overnight.
Mythos capabilities that surprised Felix
Code security auditing
Mythos appears to be the first system that truly combines the speed of static analysis tools with the accuracy of human auditors. It can find complex vulnerabilities requiring cross-file context, business-logic intent, and attack-surface judgment.
Autonomous action
Once reasoning crosses a threshold, capabilities become qualitative, not quantitative. The jailbreak-email example shows:
- Understanding implicit goals (not just jailbreak, but help researchers understand it)
- Selecting optimal communication channel (email)
- Generating structured output (detailed report)
What is Claude Cowork?
Not a simple extension of Claude Code.
Claude Cowork is designed for scenarios where AI acts as a "colleague" participating in sustained work:
- Remembers project state
- Tracks prior decisions and their rationales
- Understands codebase-specific constraints
Key distinction:普通对话是"一问一答";Cowork 是"项目伙伴"模式。
Core engineering insights from building Cowork
1. State management > model capability
The hardest problem was not generating better code, but state management:
- Where are we in the project?
- What decisions were made before?
- Why were they made?
This drove massive investment in context engineering and state tracking.
2. Controllability > raw intelligence
When users first experience Cowork, their top concern is not "is it smart enough?" but "I don't know what it's doing."
The stronger the model, the more users need to see its reasoning process. Otherwise, wrong decisions are neither understandable nor correctable.
This shaped UX priorities: transparency and intervenability are first-class.
3. Error recovery > error avoidance
Perfect AI is unrealistic. Cowork's design philosophy: assume AI will make mistakes, but make them easy to discover and fix.
Architecture supports:
- Rollback
- Branch experiments
- Compare alternative approaches
A Git-like workflow, but for AI collaboration.
Execution is Free
Felix's widely quoted phrase: "Execution is free."
Traditional constraints
Software entrepreneurship has always been constrained by execution cost:
- Idea validation needs prototypes → engineers + designers + weeks
- Product development needs more engineers + testing + iteration
- Market validation loops are expensive
The new arithmetic
When a natural language description becomes runnable code in minutes, the constraint shifts from "can we build it?" to "what should we build?"
Clarifications:
- Fast execution ≠ correct execution. Strategy and judgment rise in value.
- Prototype ≠ product. Production-grade software still requires engineering judgment.
- Code ≠ business. Finding the right idea and understanding users remain unchanged.
New entrepreneurial barriers
- Non-technical founders: May no longer be blocked by technical execution if they can express requirements clearly.
- Technical founders: Must shift from "I can write code" to "I can judge what's worth writing."
- Investors: May shift from "can this team execute?" to "does this team know what to execute?"
The end of SaaS
Felix argues that traditional SaaS is ending.
SaaS = humans buying software on behalf of agents
A decision-maker logs into Salesforce, fills forms, configures permissions, learns the UI, and recommends the tool to the team.
Agent-first = agents buying headless APIs on behalf of humans
No UI. No configuration wizard. One agent calls another agent's API to complete a task.
Example: sales data analysis
- SaaS mode: log into BI tool → upload/connect data → configure report → wait for charts → export
- Agent-first mode: AI assistant calls data analysis API → sends data + requirements → receives structured result → integrates into workflow
Headless ≠ Brainless
Headless APIs may be more complex than SaaS:
- Stronger error handling (no human sees error popups)
- More precise parameter definitions (no UI guidance)
- Better state tracking (no visual progress bars)
"Designing an interface that an agent can understand and use is harder than designing a UI for humans."
Timeline
Felix predicts 3–5 years for headless/API-first to move from edge to mainstream.
Accelerating factors:
- Agent capability leaps (Mythos-level models)
- Cost pressure (headless APIs are typically cheaper)
- Integration demand (agent-to-agent is more efficient than human-operating multiple SaaS platforms)
Skill migration: from computer language to human language
Future successful software builders will shift from "knowing computer languages" to "knowing human language."
"Human language" capability means:
- Requirements decomposition — turning fuzzy ideas into executable steps
- Boundary definition — clearly stating what's in, what's out, what's negotiable
- Validation criteria — defining "done" and quality standards
- Context provision — who are the users, what's the scenario, what are the constraints
It's not either/or
- AI can write code, but doesn't know what code to write
- AI can debug, but doesn't know correct behavior
- AI can optimize, but doesn't know what's worth optimizing
The ideal builder combines:
- Clear human-language problem description
- Enough systems knowledge to judge AI-generated solutions
Implications
- New entrants: programming syntax matters less; communication and analysis matter more
- Experienced developers: architecture, systems thinking, and quality judgment appreciate; raw typing speed depreciates
- Team structure: "technical translators" (business needs → technical needs) become more scarce and valuable
UX is harness, not model
Felix's answer to "what is good AI UX?":
"The successful AI agent is not a more intelligent model, but a better harness — one that helps users organize work and build trust."
The paradox
Model capabilities grow exponentially, but human ability to manage complexity grows linearly. A smarter model without complexity management worsens UX.
Example: "Optimize this database query"
- Bad UX: AI silently changes the code → user feels anxious
- Good UX: AI shows its thinking — "I found a missing index, considered A (faster, more memory) and B (balanced), recommend B" → user feels in control
Three harness functions
- Organize work — project stage, past decisions, next steps
- Build trust — predictable behavior, visible reasoning, intervention points, rollback
- Manage cognitive load — filter noise, highlight key decisions, hide irrelevant details
Cowork UX implementations
- /plan mode: AI shows plan before executing. User can review, modify, approve.
- Project status view: always-visible "where we are, what's done, what's pending"
- Progressive disclosure: complex tasks unfold step by step; user can drill in or accept defaults
Historical analogy: the printing press
Felix places today's AI at "Gutenberg just built the first printing press."
- The technology exists but is still primitive and expensive
- Only a small fraction has access
- No one can foresee the consequences: mass literacy, newspapers, scientific revolution, modern democracy
Near-term certainties (2–3 years)
- AI becomes standard developer infrastructure, not optional plugin
- Role boundaries blur: designers generate code, engineers generate design, PMs validate ideas directly
- New skill hierarchy: problem definition > AI guidance > direct execution (execution layer shrinks)
Long-term: humility
"Whenever someone tells me they can predict AI 10 years out, I'm skeptical. We don't even know the capability boundary 2 years out."
The Mythos email example was something Felix would have placed "at least 5 years away" a year ago.
Advice for builders
- Stay adaptable — willingness to rethink "what is programming / software / work" is a massive advantage
- Invest in judgment, not just skills — skills obsolete faster; knowing what to do appreciates
- Stay alert, don't over-predict — you don't need to forecast perfectly, but you need to react quickly to step-function changes
- Focus on harness, not just model — model capabilities commoditize; harness design differentiates
2026-09-17: Cowork merged into the main Claude conversation
来源:AI Briefing 2026-09-17 Afternoon
Anthropic announced that Cowork and chat are becoming one Claude experience. The product now chooses between a quick response and deeper agentic work from the same conversation. Claude Docs, Slides, and Design are also available in that surface, so a report, presentation, and matching visual can share context instead of moving through separate products.
The release supports this page's earlier claim that Cowork is a sustained-work mode, but changes its product boundary. Cowork is no longer presented as a separate destination. It becomes a routing decision inside the main conversation. Existing chats, projects, artifacts, connectors, and skills remain available, and users can stop, redirect, edit, export, or share generated work.
Counterpoints & Gaps
- The source is an official product announcement and builder commentary. It does not provide routing accuracy, failure recovery rates, or independent user studies.
- A unified entry point can reduce product-selection burden while also hiding which execution mode, permissions, or tools were chosen. High-risk work still needs visible routing and approval state.
Sources (addendum)
Pending Jean review
- claude-code/claude-cowork-felix-rieseberg —
ai_draft; suggested action: verify the unified surface, routing visibility, and intervention controls after rollout reaches the local account.
2026-10-28 — 取消传统编码面试与初级工程师建议
来源:AINews Anthropic ends Coding Interviews
Felix Rieseberg 以 Anthropic 工程负责人与招聘经理身份宣布本月停止传统编码面试。其理由:"给你一道 React 小题和考你汇编一样没有意义,Claude 在这两项上都比人强。我们最有影响力的构建者,无论在这些测试上好坏,都是成功的。"评估方式改为 work samples 与业务影响评价。
同文给出 "Hunger beats tenure" 主题下给初级工程师的五条建议:
- 选择业务影响而非 backlog:工程师被雇佣是来创造业务价值的,不是来写程序的(引 Patrick McKenzie 2011 年的说法)。
- 选一个激进的想法:提出少数几个一旦成功就能显著提升组织达成目标几率的想法。
- 用一个周末把它做出来:小团队或单人工程师可以用可工作的原型跳过 PRD、立项与为设计或产品资源游说的常规流程。
- 聚焦一个好主意:初级人员常犯的错误是把多个想法塞进一个原型,把三个好主意埋在 20 个坏主意里。
- 利用新人视角:做面向用户的工作时,新人对用户有独特同理心,更清楚产品对用户好用还是难用。
这与本页 "Execution is Free" 与技能迁移判断一致:当编码执行被 AI 接管,招聘与评估的轴心从"能否写代码"转向"能否判断该写什么、是否值得做"。
Counterpoints & Gaps
- 这是单一招聘经理的实践声明,不代表 Anthropic 全部 hiring bar。
- work samples 与业务影响评价的考核信度与公平性未见数据。
- "Claude 在 React 小题和汇编上都比人强"是演讲式表述,未附带评测依据。
Sources (addendum)
Related
- Claude Cowork Dispatch 发布概述(页面已合并至 claude-code/overview)
- ai-ecosystem/anthropic-mythos — Claude Mythos capabilities and data
- product-trends/saas-extinction — SaaS extinction and agent-first future
- harness-engineering/overview — Harness engineering overview
- claude-code/overview — Claude Code overview