Defining Block’s AI Agent Orchestration Strategy
Block’s AI agent orchestration strategy combines a Slack-driven control layer, a custom agent framework, and a consolidated monorepo so AI coding agents can research, plan, and change code across hundreds of services in a single, automated workflow. Rather than treating AI tools as isolated copilots inside one repository, Block built Goose, an open-source framework for running many agents, and Builderbot, a higher-level system that coordinates them. Builderbot connects to issue trackers such as Linear and Jira, then drives the full lifecycle: creating branches, opening pull requests, watching CI, and iterating on feedback without leaving a Slack thread. This multi-framework approach allows the conversation to become the development environment while Goose and Builderbot coordinate operations against a JVM monorepo that now backs thousands of weekly builds, turning AI coding agents into part of the production software development automation pipeline.
Slack as the Operations Console for AI Coding Agents
For Block, Slack is not only a messaging tool but the main console for AI agent orchestration. Engineers summon @builderbot in a channel, describe the task in natural language, and watch as the agent starts its research and planning in the same thread. Builderbot can pull work directly from Linear or Jira, tie each change to a ticket, and progress from design to pull request while humans supervise. Several teammates can follow the thread, correct assumptions, and adjust scope in real time, which reduces context switching between chat, IDE, and CI dashboards. According to Block, the system runs more than 200,000 operations a day and merges about 1,500 pull requests a week, representing roughly 15% of all production code changes. This shows that chat-native control can scale from convenience feature to core enterprise AI infrastructure.
Builderbot and Goose: Multi-Agent Orchestration Across Cash App and Square
Single-repository coding assistants broke down when Block tried to apply them to hundreds of interconnected services and hundreds of millions of lines of code. Builderbot, built on top of the Goose framework, addresses that by coordinating multiple AI coding agents that understand every internal service, API, and engineering convention. The framework gives agents the context and permissions to touch any repository, so an engineer working on Cash App can ask Builderbot to modify a Square backend service they have never used. Builderbot acts as a central orchestration layer that can handle anything from routine bug fixes to large architectural migrations spanning multiple databases. It tracks CI results, reacts to failures, and updates code until checks pass. Block describes Builderbot as “the missing layer between AI coding tools and how engineering actually works at scale,” highlighting the gap between single-model tools and enterprise AI infrastructure needs.
Monorepo Migration: Foundation for Software Development Automation at Scale
Block’s move from about 450 separate JVM repositories to a single monorepo gave Goose and Builderbot a more predictable environment for software development automation. In the old polyrepo model, shared libraries lived in their own repos and teams advanced versions at different speeds, causing dependency drift, diamond dependency surprises, and frequent runtime failures. The monorepo lets agents and humans apply atomic changes across services in one commit and resolve dependencies directly from source instead of juggling internal library versions. Block reports that the monorepo now supports roughly 8,800 builds per week with p90 CI times around 10 minutes on a reliably green main branch. As senior engineering leaders have noted, the migration became a step-change in developer experience, enabling faster IDE workflows, safer shared upgrades, and a platform where AI agent orchestration can operate on a consistent, organization-wide dependency graph.
Architectural Lessons for Enterprise AI Infrastructure
Block’s approach offers a set of architectural patterns for enterprises exploring AI agent orchestration. First, make the collaboration platform the orchestration plane: by centering operations in Slack, Builderbot keeps humans in the loop and aligned on every automated change. Second, treat AI coding agents as first-class actors in your delivery pipeline, wired into issue trackers, CI, and version control, not isolated code generators. Third, combine a multi-agent framework like Goose with a simplified code topology; the JVM monorepo cuts coordination overhead so agents can reason over a single dependency graph instead of hundreds of drifting repos. Finally, multi-agent orchestration enables tasks that a single model cannot handle, such as cross-service migrations that demand long-running planning, retries, and coordination. Together, these patterns turn AI coding agents from experimental tools into dependable enterprise AI infrastructure for large-scale software development.
