Block explains how AI coordinates thousands of code changes while humans control deployment

Block says AI can now coordinate company-wide software changes that once required humans to organize hundreds or thousands of separate contributions. In an October 8, 2026 interview published by Anthropic, AI capabilities chief Bradley Axen described Claude Fable 5 directing smaller coding models across those migrations—while engineers still control what gets merged and deployed.

One planner, dozens of coding workers

Block’s migrations can span multiple code repositories and millions of lines of code. Each pull request is a proposed set of changes for review; a migration can involve hundreds or thousands of them.

Previously, people coordinated that work while models handled one or a few pull requests at a time, Axen said. A human engineer now steers the migration, while Fable tackles higher-level decisions about data models, the rules for connecting software components, and algorithms.

Fable then directs up to dozens of smaller models, including Opus or Sonnet, to edit files and run tests. Block runs the workflow in Buzz, its open-source workspace where people and agents share channels and threads. Axen says that shared setting gives engineers visibility into how the work is progressing.

The strongest model is not the default for every task

Early results, Axen says, show Fable doing complex planning upfront, then shifting a greater share of the processing to smaller worker models without losing quality. The interview gives no numerical savings or quality measurements for that claim.

Block is building an automatic selector to help employees choose a model for each task. Its evaluations compare every model at every effort level. Axen wants clear recommendations for particular uses, with tools applying them automatically where possible and experts free to explore.

Agents can propose changes, not release them alone

Axen says agents’ changes must pass security checks. People retain control over merges into the main code branch, production deployments and related changes such as feature flags. Two humans must approve a deployment.

Other guardrails sit outside the model’s decision-making loop. Axen described inexpensive automated checks built into work interfaces, sometimes using simple text-matching rules, with smaller classifiers increasingly involved. Block also uses network allow lists to permit specified connections.

Block lets agents open pull requests, review code and query logging systems and analytical databases. Axen says Claude refuses requests to bypass dual approval in Block’s testing—not an independent demonstration of the controls.

More code also creates more upkeep. Axen says increased output is accelerating technical debt—the cleanup needed as software grows. Agents now make daily or weekly passes through codebases to identify what should be cleaned up or consolidated.