At Block, artificial intelligence (AI) is integral to our development processes. Over the last two years, we have concentrated on embedding AI into the work culture of our engineers. This includes open-sourcing Goose, our AI agent framework, collaborating with Anthropic to develop the Model Context Protocol (MCP), and creating internal tools that incorporate AI into the daily tasks of our engineers. Currently, every engineer at Block leverages AI in their workflow.
Despite these advancements, we encountered limitations. Traditional coding tools excel within single repositories but struggle with the complexity of our environment, which encompasses hundreds of millions of lines of code and numerous services. To address this challenge at scale, we introduced Builderbot.
Builderbot functions as an orchestration layer that connects various AI agents throughout our codebase. Integrated into Slack, users can interact with Builderbot by tagging @builderbot along with a brief description of their request—be it a bug fix, service migration, or feature implementation. This allows for real-time collaboration, where team members can oversee the process as Builderbot conducts research, plans, and executes tasks without the need for context switching, as discussions occur directly within the development environment.
What distinguishes Builderbot from standard coding assistants is its extensive scope. Builderbot comprehensively understands Block's entire codebase, covering all services, APIs, and conventions. This allows an engineer working on Cash App, for instance, to modify a Square service they have never directly engaged with, as Builderbot already has the requisite knowledge. It seamlessly handles tasks by picking up tickets from Linear and Jira, creating branches, writing code, and opening pull requests—all while monitoring continuous integration processes and adjusting based on feedback. Human intervention is reserved for areas where their expertise adds the most value.
Importantly, Builderbot operates solely on source code and system configurations, ensuring that it does not access or handle customer data, payment details, or personally identifiable information.
The impact of Builderbot on our development processes is significant. It executes over 200,000 operations daily and merges around 1,500 pull requests each week, accounting for approximately 15% of all production code changes at Block. What previously required months can now be accomplished in a matter of days.
"The best way to conceptualize Builderbot is as the crucial layer that bridges AI coding tools with large-scale engineering practices," explained Brad Axen, Block’s Head of AI Capabilities. "It manages the orchestration, context, and environment, allowing our engineers to focus on important problem-solving. For instance, on the Square side, we expedited the launch of features that sellers had been awaiting for months, accomplishing this in just days thanks to Builderbot handling the foundational and repetitive tasks while our engineers made product-shaping decisions. This transformation allows an idea to go from backlog to being live for millions of customers in a matter of days instead of months."
Builderbot is constructed on Goose, our open-source agent framework, which we developed and shared with the Agentic AI Foundation (AAIF) under the Linux Foundation. The challenges we faced while developing Goose inspired our partnership with Anthropic on MCP, which is now recognized as an industry standard for linking AI agents with tools and data sources.
We created Builderbot out of necessity. The issues we’re tackling are not exclusive to Block; many organizations face challenges in orchestrating AI agents across extensive codebases while ensuring quality at rapid speeds and allowing humans to concentrate on critical decisions rather than mundane tasks. We are sharing our experiences in building Builderbot because we believe the transition from AI-assisted coding to AI-native engineering is a vital discussion within the technology sector, and we aim to contribute to this conversation.
