Block uses Claude Fable 5 to orchestrate thousands of pull requests in large-scale migrations

Block has shifted its engineering approach by using Claude Fable 5 as an orchestrator for large-scale code migrations, a task that previously required significant manual human coordination. Instead of relying on a single model to handle all edits, Fable manages the high-level design—such as data models and API specs—and directs smaller, cost-effective models like Opus and Sonnet to execute individual file edits and run tests. This hierarchical orchestration allows the team to merge up to a thousand pull requests per migration while maintaining quality and reducing costs.

Orchestration and Model Efficiency

The core of Block's strategy is moving up a level of abstraction. Fable handles the complex, upfront planning using frontier tokens, while the bulk of the mechanical work is delegated to worker models. This approach marks a departure from the previous norm of using the most powerful model for every task. Block utilizes its open-source workspace, Buzz, to facilitate this multi-agent workflow, providing visibility into how human engineers and AI agents collaborate in shared channels. The result is a system where engineers focus on hard design decisions rather than repetitive code changes.

Safeguards and Human Oversight

Despite the scale of automation, Block maintains strict human-in-the-loop safeguards for critical operations. Agents are permitted to open pull requests, review code, and query databases, but they cannot merge to main or deploy to production without approval. The company enforces a dual-approval layer for deployments, ensuring that no LLM can push changes to production independently. Additionally, Block employs a network allow list and efficient, non-LLM security checks (such as regex and small classifiers) to prevent common errors and security risks, keeping the model loop focused on productive tasks while relying on Anthropic's safety classifiers to refuse dangerous requests.