Mike Krieger on Building with Claude Fable 5
the gist
Mike Krieger, co-founder of Instagram and head of Anthropic Labs, discusses how the high-reasoning capabilities of Claude Fable 5 have fundamentally shifted his workflow from manual coding to delegating complex, multi-step architectural tasks to an autonomous agent.
The Shift to Agentic Delegation
Mike Krieger describes a fundamental evolution in his personal workflow: moving from being a hands-on coder to an architect who delegates complex, long-running tasks to Claude Fable 5. He notes that the model's ability to maintain global context and handle multi-step reasoning allows him to offload entire projects, such as building a media tracker, to the model. He often sets up complex tasks before going to sleep or heading out, returning to find the work completed, documented, and scaffolded, even if the model encountered minor service interruptions.
Architectural Planning and Human-in-the-Loop
Despite the model's autonomy, Krieger emphasizes that the human role has shifted toward high-level architectural planning and verification. He uses the model to generate diagrams, markdown documents, and initial prototypes to align his team before execution. He highlights the importance of "verification" as a critical step in the development process, noting that while the model can execute, the human must still validate the output. He also discusses the necessity of maintaining multiple concurrent sessions—some for high-context, long-running tasks and others for rapid, iterative feedback.
Democratizing Software Creation
Krieger reflects on the collapse of the cost and time required to build software. Comparing his experience building Instagram v1—which required days of all-nighters—to his current ability to build functional apps over a weekend, he argues that the gap between intent and execution has narrowed significantly. He shares an anecdote about a non-technical recruiter at Anthropic who used these tools to build her own internal software, illustrating that these models are empowering a new class of builders who previously lacked the technical gatekeeping skills to bring their ideas to life.
Agent-Native Architecture
Krieger advocates for "agent-native" software design, where applications are built with the assumption that an AI agent will be interacting with them. This includes making every feature accessible via tool calls and enabling the agent to modify the software from within itself. He demonstrates a personal project where he can long-press a chat interface to trigger agentic edits to the app's UI, effectively closing the loop between user feedback and code deployment.