Why

AI agents perform well on greenfield projects (new codebases built from scratch) but struggle with existing brownfield codebases (mature, production systems with existing code).

The core problem: they burn most of their context window on exploration—searching files, tracing flows, connecting dots—leaving little capacity for actual implementation. By the time they're ready to code, they've hit the "dumb zone" where performance degrades sharply. And since they lack long-term memory, this cycle repeats with every conversation.

Multi-repo Complexity: The problem compounds with multi-repo architectures. When code is split across connected repositories, the agent exhausts its context just mapping dependencies between codebases—often before writing a single line.

Internal Dependencies: Internal dependencies and niche packages present another failure mode. The agent has no onboarding to proprietary systems, so it hallucinates usage patterns. Worse, when internal documentation has drifted from the actual implementation, the agent trusts those "lies" and produces code that doesn't work.This problem compounds dramatically when working outside the most popular ecosystems (Python and TypeScript/JavaScript).

The end result is the same across all scenarios: slop code requiring heavy rework.