AI coding tools are starting to diverge on a technical question that matters to developers paying for tokens and reviewing machine-written patches: should an agent build a map of a codebase before it starts work, or should it search as it goes?
Augment Code is arguing for the map. Vinay Perneti, the company’s vice president of engineering, told Ars Technica that Augment’s product pre-indexes a repository using embeddings, a retrieval model and a vector database. When a developer assigns work to an agent, the system retrieves code that appears conceptually relevant rather than relying mainly on keyword search.
That is a different bet from Anthropic’s Claude Code team. Cat Wu, Anthropic’s head of product for Claude Code, previously told Ars Technica that Anthropic favors a leaner harness because model capabilities are changing quickly. She said Anthropic had not seen measurable evaluation gains from adding more opinionated code-navigation tools by default.
A harness is the application layer around a large language model. It decides what the model can inspect, which tools it can call, how it edits files and how it interacts with the developer’s project. The model may get the attention, but the harness determines how much useful code reaches the model before it starts guessing.
Augment’s bet: retrieval beats repeated searching
Perneti said Augment’s approach has two main parts: an embedding and retrieval model pair, plus a vector database and backend built to return relevant results quickly. He described the system as aimed at large private repositories, where the model has not already absorbed the code during training.
His argument is straightforward enough: public benchmark repositories may already be familiar to frontier models, so an agent can appear to know where to look. Private codebases are a different problem. The model has to discover internal conventions, duplicated patterns and hidden dependencies during the task, unless another system gives it that context.
Perneti said Augment’s own testing showed a token-efficiency benefit. In a company blog post comparing Augment Code and Claude Code on Terminal-Bench using the same model, he said Augment reached similar accuracy while using 33 percent fewer tokens. That is Augment’s claim, not an independent audit, and benchmark results in AI coding remain easy to overread.
He also pushed back on the idea that failed retrieval experiments settle the issue. Perneti said retrieval systems vary in quality, just as databases do, and said Augment spent about 18 months after its 2022 founding researching embedding and retrieval models for large codebases.
Cost, trust and technical debt
Perneti did not claim agents are ready to replace engineering judgment. He told Ars Technica that handing work to agents and walking away is the wrong pattern. In his view, humans still need to shape specs and review decisions, while agents are better at executing against a well-defined plan.
He also acknowledged a failure mode developers already complain about: agents duplicate code. Perneti said Augment has seen that internally and has run focused efforts to reduce technical debt using agents, with humans defining what cleanup should mean.
On cost, Perneti said companies should think about how much of their token budget is being spent on gathering context. He also suggested that teams may eventually reserve expensive frontier models for harder problems and use open-source models for routine, well-specified coding work.
The disagreement is narrower than it may sound. Anthropic and Augment are not necessarily evaluating the same tools or optimizing for the same outcomes. Wu discussed limited code-navigation additions such as language server protocol tools, while Augment is describing a broader retrieval system built around pre-indexed repositories.
The useful takeaway for developers is less glamorous than the marketing: coding agents are not just models in a chat box. The surrounding software decides whether the agent searches blindly, receives curated context or burns tokens rediscovering the obvious. That layer may decide whether AI coding feels like leverage or an expensive autocomplete with a commit bit.
This story draws on original reporting from Ars Technica.
