Two competing philosophies are emerging in AI-assisted coding: lean harnesses that trust model improvements versus context-rich systems that pre-index codebases.

Anthropic's Claude Code represents the lean camp. Product head Cat Wu told Ars Technica the team avoids building structured context around codebases by default, seeing no measurable evaluation gains from semantic code-navigation tools.

Augment Code takes the opposite approach. The startup pre-indexes repositories using embeddings, retrieval models, and vector databases to surface conceptually relevant code before models begin work.

The semantic advantage

Vinay Perneti, Augment Code's VP of Engineering, argues the benefits appear in large private codebases where models lack prior training exposure.

"For all the public, open source repos where most benchmarks are run, every single model has basically memorized the repo," Perneti said. "Whereas when you're doing this in a private repo, a model has never seen that repo."

Augment Code claims its semantic approach delivered 33% better token efficiency than Claude Code on Terminal-Bench evaluations while maintaining similar accuracy.

The company spent 18 months researching retrieval and embedding models specifically for large codebases, starting in 2022 before ChatGPT's release.

Intelligence versus context

Perneti acknowledges models are improving rapidly but argues intelligence alone isn't sufficient.

"There's two ingredients that you need for higher quality outcomes, intelligence and context," he said. "Just because they're more intelligent does not mean they have the context."

The debate reflects broader questions about AI development infrastructure. Cursor and other coding tools face similar architectural decisions about how much structure to impose versus letting models handle context discovery.

Augment Code's retrieval system operates in sub-milliseconds, suggesting the performance overhead of semantic indexing may be manageable as models scale.

The company's approach assumes context engines will remain valuable even as foundation models improve, betting that structured knowledge retrieval will complement rather than be replaced by raw model capabilities.