Source: ai-research/claude-blog-block-claude-fable-prs-2026-10-08.md
Product: Claude Enterprise (Fable 5)
URL: https://claude.com/resources/articles/how-block-orchestrates-claude-fable-across-thousands-of-pull-requests
Published: 2026-10-08
Author: Bradley Axen, Head of AI Capabilities at Block (interviewed by Anthropic)
Block (the parent of Square and Cash App) uses Claude Fable 5 for company-wide code migrations — hundreds to thousands of pull requests at a time across multiple repositories and millions of lines of code. Their open-source workspace, Buzz, gives human and AI agents shared channels and threads for collaborative progress visibility.
Key Takeaways
- Fable as orchestrator, smaller models as workers. Fable handles high-level design (data models, API specs, algorithms). Up to dozens of Opus or Sonnet instances do the file edits and test runs. No quality drop from this split.
- One migration merged ~1,000 pull requests. Engineers focus on hard design decisions, not mechanical changes.
- Benchmark models AND effort levels together. A lower-effort Fable may outperform a higher-effort alternative. Separate benchmarks for each dimension miss this interaction.
- Human approval on all merges and deploys. Two humans must approve any production deploy. Model-level protections remain on — Fable refused Block’s own tests to bypass dual-approval.
- Daily/weekly automated passes across ~1M lines to find cleanup and limit tech debt growth.
Orchestration architecture
| Layer | Model | Role |
|---|---|---|
| Orchestrator | Claude Fable 5 | High-level design — data models, API specs, algorithms |
| Workers | Opus or Sonnet (up to dozens) | File edits, test runs, mechanical changes |
Token economics: frontier tokens go to upfront planning; a higher share of total tokens go to cheaper worker models. Block reports no quality drop from this split.
Scale numbers
- One migration: ~1,000 pull requests merged
- Scope: hundreds to thousands of PRs across multiple repositories
- Ongoing: daily/weekly automated passes across ~1M lines of code
Safeguards
- Human approval required for: merges to main, production deploys, feature flags, security checks.
- Two humans must approve any production deploy — Fable’s built-in protections refused attempts to bypass this in Block’s own tests.
- Secure-by-default infrastructure: remote workstations and network allow-lists (not deny-lists).
- Cheap automated checks (regex, small classifiers) built into employee interfaces, outside the model.
Model selection and cost
- Full frontier model access for technical staff.
- Building an auto-selector to recommend models per task type — preventing wasteful use (e.g., Fable for trivial edits).
- Each use case is benchmarked across models AND effort levels. A lower-effort Fable sometimes outperforms a higher-effort Sonnet — the optimal pairing isn’t predictable without measurement.
Practical patterns (from Block’s experience)
- Use frontier models for planning and orchestration; smaller models for execution.
- Benchmark models and effort levels together, not separately.
- Keep human approval on merges and deploys, enforced with cheap automated checks outside the model.
- Capture institutional knowledge at the system level (not with individuals) so agents have best-available context.
Buzz — open-source workspace
Buzz is Block’s open-source workspace where humans and AI agents collaborate in shared channels and threads. It provides team-wide visibility into migration progress across hundreds or thousands of concurrent PRs.
Related
- Claude Code Review GA — the complementary review layer for agent-generated PRs at scale.
- Claude Code Agent Teams — the local multi-agent primitive; Block’s Fable+workers pattern is an enterprise analog.
- Agent Observability on the Claude Platform — OTel tracing surfaces per-step latency and token cost across the orchestrator+worker pipeline.
- Claude Code Subagents — the subagent primitive underlying the worker layer.
- Claude Managed Agents — the hosted-agent runtime for workflows that don’t run inside Claude Code.