AI Coding Agents Are Breaking Big Codebases — Dan Adler, Sourcegraph
The talk addresses the critical challenge of maintaining massive, decades-old codebases (tens of thousands of repositories) in the era of AI coding agents. While agents accelerate development, they also accelerate codebase decay through duplicated code, drifting standards, and new vulnerabilities. Dan Adler argues that the bottleneck is not model quality, but the infrastructure required for large-scale code visibility. Sourcegraph introduces 'Agentic Batch Changes,' a product designed to execute code changes across thousands of repositories from a single prompt, ensuring auditability and consistency across complex, distributed systems.
Key takeaways
-
The Scale Problem
Most software industry employment is in large enterprises managing thousands of repositories and decades of code history, making codebase ownership increasingly difficult.
-
Codebase Decay
4:17
AI-generated code, while fast, is causing decay by proliferating duplicated code, creating brittle cross-service dependencies, and introducing new vulnerabilities.
-
Visibility is Infrastructure
The core problem is the volume of code; agents cannot understand what they cannot see. Large-scale code visibility and a comprehensive 'code graph' are essential infrastructure requirements.
-
Agentic Batch Changes
Sourcegraph launched 'Agentic Batch Changes,' a frontier agent that allows owners to execute code changes across hundreds or thousands of repositories from a single prompt, providing necessary tracking and auditability.