What Block’s AI Agent Orchestration System Does
Block’s AI agent orchestration system is an internal platform that coordinates multiple AI coding agents, connected through Slack, to perform research, planning, code changes and CI monitoring across thousands of services in a unified way. Built on the open-source Goose framework and branded internally as Builderbot, the system focuses on AI code generation that can cross repository boundaries and handle large, interconnected codebases. Instead of a single assistant tied to one repository, Block treats AI coding agents as a fleet that can be steered in real time from a chat thread. The goal is software development automation that fits how engineers already work: tickets come from tools like Linear or Jira, Builderbot creates branches and pull requests, and the conversation becomes the development environment rather than a separate IDE.
Builderbot and Goose: AI Coding Agents from Slack
Builderbot sits on top of Goose as a central agent orchestration system, coordinating multiple AI coding agents that understand Block’s codebase, APIs and engineering conventions. Engineers summon it by tagging @builderbot in Slack and describing the change: from bug fixes to sweeping architectural tasks. Builderbot then picks up issues from tracking tools, creates branches, opens pull requests, watches CI, and iterates on feedback while staying inside the Slack thread. Several developers can watch and guide its behaviour in real time, steering the plan and correcting missteps. According to Block, the system runs more than 200,000 operations a day and merges about 1,500 pull requests a week, covering roughly 15% of production code changes. This scale pushes AI code generation beyond mere autocomplete and into a continuous, multi-agent workflow that fits enterprise engineering.
Cross-Service Development and Monorepo Architecture
A key design goal for Block was software development automation that works across hundreds of services and hundreds of millions of lines of code. Builderbot maps the structure of the entire corporate codebase, cataloguing services and APIs so it can perform cross-service changes. An engineer working on Cash App can safely modify a Square backend they have never touched because the agents supply the missing architectural context. To reduce coordination friction, Block migrated about 450 JVM repositories into a single monorepo, cutting dependency drift and duplicated upgrade work. The monorepo now supports around 8,800 weekly builds with p90 CI times near 10 minutes on a stable main branch. This monorepo architecture pairs neatly with Builderbot: atomic updates across services in one commit, with shared dependencies resolved directly from source instead of scattered versioned libraries.
Scaling Enterprise Automation While Keeping Humans in Control
Block’s setup shows how an AI code generation stack can move from helper to infrastructure while keeping engineers in control. Builderbot acts as a central orchestration layer that owns the mechanics of cross-system execution—branching, editing, testing, and monitoring—while humans stay in the Slack conversation to set goals and approve changes. The company reports that work that once took months can now complete in days, and engineers pulled long-requested Square seller features from backlogs and shipped them in a matter of days by letting Builderbot handle scaffolding. At the same time, Block confines the system to source code and configuration, avoiding customer or payment data and ensuring each agent only has the permissions it needs. The result is a model where AI coding agents and a monorepo architecture combine to support thousands of builds a week without losing transparency or oversight.
