Mike Krieger explains how stronger models changed his own work from step-by-step task delegationAI agent delegation assigns a bounded task, context, authority, and expected result from one participant to an agent or another agent. toward describing an end state and letting an AI agentAn AI agent is a system that observes context, decides what to do, and takes actions through tools to pursue a goal. develop, test and revise a solutionIterative AI code generation repeatedly creates, runs, evaluates, and revises code instead of treating the first output as final.. He gives the example of asking Claude to port a large Python project to TypeScript over a weekend, including verification and repeated improvement of its own output.
Inside Anthropic, Mike Krieger says asynchronous delegationAn asynchronous AI agent performs delegated work in the background so a person can continue other work and review the result later. increasingly happens through shared tools where teams can see how colleagues use Claude. That multiplayer visibility encourages more ambitious requests, but also moves the bottleneck from producing code to understanding intent, architecture and tradeoffs. Large changes therefore travel with explanatory artifacts, while humans use Claude to investigate review questions instead of reading every generated line manually.
Anthropic Labs organizes work around two-week persevere-or-pivot reviews rather than permanent project teams. People assemble around a bet, and projects that gain traction become structured teams. Mike Krieger also argues that vertical AI remains open to startups with deep user knowledge, that regulated uses need flexible agents grounded in verifiable dataA grounded AI agent bases its answers and actions on relevant, verifiable information rather than relying only on model memory or unsupported inference., and that rapid AI development still requires deliberate time off and honest discussion of stress.
Watch on YouTube
