Block Merged About a Thousand Fable-Orchestrated Pull Requests, With Humans Approving Every Merge to Main

Block has merged about a thousand pull requests orchestrated by Claude Fable in a single migration, and it still requires a human to approve every merge to main. Bradley Axen, Block’s Head of AI Capabilities, described the setup in an Anthropic interview: Fable handles high-level design such as data models, API specs and algorithms, then directs up to dozens of smaller models like Opus and Sonnet that edit files and run tests while a human engineer steers.

The cost logic is the part worth copying. Early results suggest that spending frontier tokens on upfront planning and shifting most of the volume to smaller workers keeps quality while lowering cost. Axen treats model and effort level as a grid to optimize per use case, and Block is building an auto-selector. His read: “Fable on low might just be better than another model on xhigh.” The worker tier is getting cheaper too — Anthropic’s Claude Haiku 5.5 is priced at $0.10 per million input tokens for requests up to 100k tokens, aimed at exactly this kind of high-volume work.

The governance is what makes it shippable. Block requires human approval for merges to main and production deploys, two-person approval on deploys, security checks, and cheap non-model checks. Axen’s own summary: “we’re not ready to let LLMs loose on our codebase.” Agents also make daily or weekly passes over a roughly million-line codebase to find cleanup work.

For consulting and technical leaders, the outcome to wield is a delivery pattern: one expensive planner, many cheap workers, deterministic checks, and a person holding the merge button. Ship the gates first. The model mix is a tuning problem you can solve later.