AI agents are flooding large codebases with code faster than teams can

The core problem. Dan Adler, CEO of Sourcegraph, argues that AI coding agents are making large-scale codebase ownership harder, not easier. The thesis is simple: agents generate code faster than humans can review, leading to duplicated logic, inconsistent standards, hidden vulnerabilities, and brittle cross-service dependencies. The result is accelerating technical debt in the very systems that run global finance, logistics, and infrastructure.

The scale mismatch. 72% of software employment is in companies of 500+ people. These organizations manage thousands of repositories and decades of accumulated code. Adler points out that today's agent tools—Cursor, Claude Code, OpenAI—are optimized for small-scale, greenfield tasks. They cannot see, search, or reason across a 50,000-repo codebase. The constraint is context window size and the lack of a unified code graph that spans an entire organization.

The concrete symptoms. Adler cites a tech leader at a top-10 car manufacturer who overheard a developer say, "I don't know what this code does. AI wrote it for me." Another leader at a major US bank, managing 90,000 repos, struggled to apply an npm security patch across every affected service. These are not edge cases—they are the emerging normal for large enterprises where agents operate in silos without organizational visibility.

The missing infrastructure. Adler's central claim is that visibility is infrastructure. Today's agent harnesses rely on .md files and limited RAG. They cannot produce compiler-accurate analysis of dependencies, let alone map runtime relationships across services. Sourcegraph's answer is a code graph: a foundational layer that gives agents search, dependency mapping, and batch change execution across arbitrary numbers of repositories.

The tradeoff and the hype. There is real signal here: large organizations will face a code-volume crisis if agent deployment outpaces codebase governance. However, the talk doubles as a Sourcegraph product pitch, and Adler's proposed solution—batch change agents that use deterministic scripts where possible—is a plausible but unproven approach. The evidence is anecdotal: two customer quotes, no benchmarks, no independent validation.

What engineering leaders should do. The takeaway is not to buy Sourcegraph immediately. It is to measure your own agent-generated code growth, establish code review baselines, and invest in cross-repo visibility before your teams lose the ability to reason about the system as a whole. The real enemy is not agents—it is the absence of infrastructure to manage what agents produce at scale.