Mike Krieger Shifts to IC Role at Anthropic

Mike Krieger, who served as Chief Product Officer at Anthropic for two years, observed a growing trend of individuals engaging directly with AI models, leading to a strong sense of FOMO (fear of missing out). He noticed that many former Chief Technology Officers (CTOs) were returning to Individual Contributor (IC) roles to get hands-on experience with AI development. This realization prompted Krieger to shift to an IC role himself, coinciding with the internal development of snapshots that later evolved into Mythos and Fable.

Anthropic is moving away from the traditional engineering paradigm where tasks are broken down into granular, human-led steps. Instead, the focus is on describing the desired end goal and allowing the AI model to execute the work, surfacing trade-offs and managing complex architectural decisions independently. Krieger frequently asks the AI to simplify intricate explanations to a more digestible level, highlighting the shift towards AI-driven execution and problem-solving.

Mike Krieger initiated a significant project to port hundreds of thousands of lines of code from Python to TypeScript for a major labs project, aiming to leverage a better deployment story with Bun. He noted that such a large-scale migration would have been considered a 'dumb idea' in past engineering contexts, even as recently as the early 2020s. However, the current capabilities of AI models make such extensive architectural refactoring feasible.

Mike Krieger successfully utilized Claude to port an entire Python codebase to TypeScript over a single weekend. The process involved continuous iteration and verification of the generated code. He described a workflow where he would 'churn and churn and churn' through the model's output, eventually returning on Monday to a complete, ported version of the project. This highlights the practical application of large language models (LLMs) in accelerating significant codebase transitions.

Krieger emphasized that defining these boundaries is often the most difficult part of the process. Prioritizing incremental steps is essential to avoid the pitfalls of trying to 'boil the whole ocean,' leading to more manageable and effective system swaps.

Krieger noted that this approach reflects how a significant portion of modern code is written, where applications are shipped with thousands of configuration flags rather than as a single, static entity. This method allows for greater flexibility and parallel development.

The use of tagging for delegation is becoming a core interaction model, enabling a multiplayer collaboration experience reminiscent of Midjourney, where users can observe and learn from how others are using the system. This method moves beyond simple requests, allowing users to assign Claude responsibilities for monitoring the codebase. This asynchronous approach reduces the need for constant, high-bandwidth back-and-forth communication.

Anthropic developed Claude Code Artifacts to improve communication by clearly articulating the purpose and intentions behind code changes. The team is shifting its focus from line-by-line code reviews to discussions about the trade-offs made and the outcomes measured in production. This approach emphasizes understanding the 'why' behind changes rather than just the 'what.'

This structured approach views project shutdowns as design successes rather than personal failures, fostering a culture of rapid iteration. This systematic review process helps prevent the need for continuous, chaotic organizational restructuring.

Managers primarily serve as supporters, ensuring alignment with individual interests. Projects begin as ad hoc groups and only solidify into more defined structures once they demonstrate momentum and traction, allowing for greater flexibility.

Recently, Anthropic removed 'Styles' to replace them with more effective 'Skills.' This approach emphasizes the necessity of evolving product primitives, supplanting older AI generation methods with newer, more efficient workflows to maintain product clarity and effectiveness.

Current distinctions between various AI interfaces, such as code, co-work, and chat, are proving confusing for users. These disparate surfaces lack interoperability and the ability to delegate tasks to one another, creating friction. The ultimate goal is to eliminate manual barriers, such as the need to copy-paste code blocks between different windows, by merging these fragmented workflows into a more unified and seamless experience.

Mike Krieger acknowledged this concern, stating that he joined Anthropic specifically to explore how AI could accelerate, rather than destroy, the startup ecosystem. This highlights the dual nature of AI as both an immense opportunity and a significant threat to established business models.

Krieger emphasized that understanding people's needs, figuring out how to reach them, actively listening, and rapidly iterating on feedback are still the hardest aspects. He also noted that small teams focused on specific problems retain an inherent advantage in speed and agility over larger, more complex organizations, underscoring that human insight and adaptability continue to be critical for business success.

In the finance sector, there's a delicate balance to strike between implementing flexible agentic workflows and maintaining strict verifiability, audit logging, and data provenance. Krieger explained that the art lies in defining the right 'cutline'—a point where these elements coexist without overly constraining the types of applications that can be built. He noted that existing legacy systems are often too rigid, creating opportunities for innovation at both the foundational and application levels of the technology stack.

Krieger acknowledged that this environment is characterized by a constant stream of new model launches and rapidly changing regulations. To keep up, weekly team meetings at Anthropic now include dedicated 'weekend AI' segments, where team members share the latest updates and developments from the field, highlighting the extraordinary speed of innovation.

Krieger advocates for adopting an athlete's mindset to navigate the extreme highs and lows of the AI industry, advising against internalizing sentiments like 'we're so back' or 'it's so over.' He referenced Ben Horowitz's concept of the 'we're effed' feeling common in startups, emphasizing that building a company is a long game, not defined by any single model or product launch. This perspective helps in maintaining emotional stability and long-term focus amidst rapid changes.

Verbalizing emotions, such as frustration or sadness, within a team creates a crucial space for collective vulnerability and honesty. Krieger noted that if one person is feeling something, it is often the case that others on the team are experiencing similar emotions. Opening up allows teams to move beyond individual emotional processing to collective problem-solving, fostering a more supportive and effective work environment.