AI coding tools split over how much context agents need
Augment Code says pre-indexing private repositories can cut token use, while Anthropic favors a leaner harness around fast-improving models.
By Maya Lindqvist · Senior Technology Correspondent
3 min read
AI coding companies are taking different paths on a core design question: whether agents should map a codebase before work begins or search for context as each task demands it. The choice matters because agentic development tools are spreading while teams weigh cost, reliability and technical debt.
Ars Technica reported that Anthropic’s Claude Code team favors what product lead Cat Wu called a lean harness: software around the model that stays relatively unopinionated as models improve quickly. A harness controls what an AI model can see, what actions it can take and how it works with a developer’s project.
Wu told Ars Technica that Anthropic has not seen measurable gains from approaches that add more structured code-navigation tools by default. She said Claude Code generally avoids building too many assumptions into the product and lets developers add tools if they want them.
Augment Code has taken another route. Vinay Perneti, the company’s vice president of engineering, told Ars Technica that Augment pre-indexes repositories using embeddings, a retrieval model and a vector database, then pulls in code that appears relevant to the task.
Private codebases are the test case
Perneti said Augment’s approach is most useful in large private repositories. He argued that public open source projects used in many benchmarks may already be familiar to major models, while private company code is unseen and harder for an agent to understand without help.
According to Perneti, semantic retrieval can shorten the search process inside private repositories by giving the agent a more direct route to relevant files and concepts. He said Augment’s system combines embedding and retrieval models with a back-end built for very fast lookups.
Perneti also pointed to Augment’s own Terminal-Bench comparison using the same model with Claude Code and Augment Code. He said Augment completed tasks with similar accuracy while using 33% fewer tokens than Claude Code, which he attributed to less time spent gathering context.
The companies may also be measuring different things. Ars Technica noted that Wu’s comments addressed language server protocol tools, while Augment describes a broader context engine built around repository indexing, embeddings and retrieval.
Cost, trust and technical debt
Perneti told Ars Technica that better models do not remove the need for context. In his view, coding outcomes depend on both model intelligence and the information supplied to the model, with token cost becoming a major concern for engineering leaders.
He also addressed developer skepticism about AI coding agents. Perneti said teams should not treat agents as systems that can be assigned work without oversight. He said humans remain better suited for judgment-heavy steps such as reviewing specifications, while agents can be effective once a clear specification exists.
Perneti acknowledged that AI tools can add technical debt, saying Augment has seen agents duplicate code internally. He said the company has used focused work with agents to reduce that debt after defining what cleanup should mean.
On cost, Perneti said teams may eventually reserve frontier models for the hardest work and use open source or smaller models for well-defined coding tasks. Ars Technica reported that some open-weight models are approaching recent frontier performance on some coding tasks, including models small enough to run locally or on in-house hardware.
Both Wu and Perneti told Ars Technica they expect frontier models to keep improving rapidly over the next year. Their disagreement is less about whether AI will change software development and more about how much structure coding tools should add around the model while that change continues.
This story draws on original reporting from Ars Technica.