What AI Agent Orchestration Means for Enterprise Codebases
AI agent orchestration is the practice of coordinating multiple specialized AI coding agents to research, plan, and modify software systems across many services, while keeping engineers in control and inside their normal tools. Block built this kind of orchestration to go beyond single-repository coding assistants that fail in environments with hundreds of services and hundreds of millions of lines of code. Instead of dropping an AI plugin into an editor, Block created Builderbot on top of its open-source Goose framework as a central “missing layer” between AI models and how engineering teams work at scale. Builderbot does not live in a single IDE; it understands company-wide services, APIs, and conventions, and it turns a Slack conversation into the development environment for complex software development automation tasks.
Builderbot as a Cross-Service Development Orchestrator
Block deployed the Builderbot agent framework across its codebase to automate complex cross-service software development tasks that generic tools could not handle. Builderbot acts as a central orchestration layer that coordinates multiple AI coding agents to work across many repositories and services. It maps the structural context of the entire codebase, cataloguing internal services, APIs, and engineering conventions so it can change almost any repository. That design means a Cash App engineer can trigger a safe change in a Square backend service without prior familiarity, because Builderbot supplies the architectural context. The framework covers the full lifecycle: it picks up tickets from Linear and Jira, creates branches, generates code, opens pull requests, and monitors continuous integration pipelines, iterating on failures and feedback. According to Brad Axen, Block’s head of AI capabilities, “Builderbot is the missing layer between AI coding tools and how engineering actually works at scale.”
Running a Fleet of AI Coding Agents from Slack
Instead of building a new interface, Block manages its fleet of AI coding agents from Slack using Goose and Builderbot. Engineers tag @builderbot in a Slack thread with a short description of the task, from small bug fixes to multi-database migrations. The agent answers and works directly in that same thread, carrying out research, planning, and coding steps in the open. The conversation becomes the development environment, so developers no longer switch contexts between chat, IDE, and dashboards. Multiple teammates can watch a ticket unfold in real time, steering the agent’s direction or clarifying requirements as it progresses. This Slack integration turns agent framework deployment into an everyday practice: all Block engineers now use AI in their normal routines, and the orchestration system fits naturally into existing communication and collaboration habits while keeping humans firmly in the loop.
From Generic Tools to Custom Orchestration Systems
Block’s experience shows why enterprises are moving beyond off-the-shelf AI coding assistants toward custom AI agent orchestration. Single-repository tools could not process the mechanical complexity of a mature environment with interconnected services and strict conventions. Builderbot, built on Goose, provides a tailored orchestration layer that understands those constraints and can coordinate many agents across the full development pipeline. This approach brings software development automation closer to real engineering work: the system retrieves tickets, edits code where needed, manages pull requests, and responds to CI and human feedback. At the same time, Block keeps strict boundaries around sensitive information and limits agents to source code and configuration data. The result is an internal platform where AI coding agents handle the repetitive, cross-service work while engineers make design choices, setting a pattern for how other enterprises can design their own orchestration systems.
Scale, Security, and Human Control in Agent-Driven Workflows
Block’s deployment highlights what large-scale AI coding agents look like in production. The system now 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. These volumes show that AI agent orchestration is not a side experiment but a core part of Block’s engineering pipeline, compressing work that once took months into days. At the same time, Builderbot operates only on source code and system configurations, with no access to customer data or payment information, so software changes stay separated from sensitive production records. Slack-centric oversight keeps humans embedded in the loop: engineers watch Builderbot’s steps, guide tradeoffs, and approve outcomes. For enterprises, Block’s model illustrates how agent framework deployment can balance aggressive automation with operational visibility, security, and clear human authority over released changes.
