The Shift to Unreasonable Delegation

Mike Krieger argues that the primary constraint on AI-powered development is a psychological one: we are still conditioned by the limitations of early AI tools. To move faster, builders must stop treating LLMs as glorified autocomplete and start treating them as autonomous teammates. This means moving from task-based delegation to 'end-state' delegation—describing the desired outcome and allowing the model to handle the implementation, trade-offs, and verification. A prime example of this is Krieger’s weekend experiment where he had Claude port a massive Python codebase to TypeScript, verify the results, and iterate until it was production-ready, a task that would have previously required weeks of human engineering.

The 'Persevere or Pivot' Org Structure

Anthropic Labs operates on a two-week cadence designed to embrace failure. Projects are reviewed every 14 days with a binary choice: persevere or pivot. This structure is intentional; it prevents the organization from becoming attached to failing ideas. To maintain this agility, the organizational chart is decoupled from project teams. 'Bet leads' are assigned to projects, but they often manage no direct reports. This allows teams to be assembled and disbanded fluidly without the overhead of constant reorgs. Once a project proves its value—like Claude Design—it is then transitioned into a more permanent, structured team.

Solving the Review Bottleneck

As AI agents generate more code, traditional line-by-line code review becomes a bottleneck. Krieger suggests that the real constraint is the human ability to conceptualize the intent behind massive changes. Instead of reviewing raw code, Anthropic now uses 'artifacts'—documents that outline the intent, the trade-offs made, and the reasoning behind the implementation. This shifts the review process from checking syntax to verifying logic and strategy. Krieger notes that he no longer reads every line of code, relying instead on Claude to investigate specific questions or concerns within the PR.

Scaling and Product Philosophy

Drawing from his experience at Instagram, Krieger emphasizes that modern AI products must prioritize observability and dynamic configuration. He advocates for 'pre-measuring' everything and building robust feature flags to allow for safe, incremental rollouts. Regarding product bloat, he advocates for 'Project Unhip'—a deliberate effort to identify and remove low-usage features. He believes that the current AI landscape is too fragmented, with separate surfaces for chat, code, and design that do not interoperate. The next generation of products will succeed by blurring these lines and allowing models to act across different environments seamlessly.

Key Takeaways

  • Be Unreasonable: Stop asking AI to do small tasks. Define the end state and let the model handle the 'how.'
  • Delegate Intent, Not Tasks: Treat the model as a teammate with memory and context, not just a tool for code generation.
  • Optimize for Comprehension: When reviewing AI-generated work, prioritize documentation of intent and trade-offs over line-by-line code inspection.
  • Embrace Rapid Failure: Use a 'persevere or pivot' review cycle to kill projects quickly and avoid the sunk-cost fallacy.
  • Decouple Teams from Projects: Keep your org structure flexible by assigning leads to projects without requiring them to manage the project's contributors permanently.
  • Unship Often: Regularly audit your product for low-usage features and remove them to reduce complexity.

Notable Quotes

  • "I think as an industry or even as a product team we have to teach people to be more unreasonable in their usage."
  • "The bottleneck I care about is not review time but whether a human can still hold the change in their head."
  • "I don't review when I get a pull request... I actually talk to Claude about the code and say, 'All right, like these are the questions that I would have, can you go investigate it?'"
  • "We shut down projects basically every single one of those cycles and it's like the more you do it the less it's just like 'oh no my project has shut down, I failed.' It's like, no, that is definitely the intention."