AI coding tools split over context-rich harnesses

The latest argument in AI-assisted software development is no longer just about which model is best. It is increasingly about the harness around the model: the software layer that decides what the model sees, what tools it can use, and how it interacts with a codebase.

That debate is now playing out between Anthropic’s Claude Code and Augment Code. Anthropic has argued for a lean harness, with fewer built-in assumptions and less structured context gathering by default. Vinay Perneti, Augment Code’s VP of Engineering, told Ars Technica his company is making the opposite bet: that richer context infrastructure materially improves coding performance, especially in large private repositories.

According to Perneti, Augment pre-indexes repositories using embeddings, a retrieval model, and a vector database, then surfaces conceptually relevant code in sub-millisecond time. He contrasted that with more grep-based approaches used by Claude Code, Codex, and other agents.

Perneti said the difference matters most outside public benchmarks. In open-source repos, he argued, leading models have often effectively already seen enough of the code to perform well. Private codebases are different because the model has never encountered them before, making retrieval and context gathering more important.

Recommended reading

Open models tighten the screws on Anthropic

He also tied the approach to token usage. Perneti pointed to an Augment blog post claiming that, on Terminal-Bench using the same model as Claude Code, Augment delivered similar accuracy while being 33 percent more efficient. He suggested that some disagreements with Anthropic may come down to what exactly is being tested, noting that not all retrieval systems are equivalent and that Augment spent about 18 months after the company was founded in 2022—before ChatGPT—researching retrieval and embedding models for large codebases.

Context, cost, and trust

Perneti framed the problem as a balance between intelligence and context. Models will keep improving, he said, but that does not mean they automatically have the right project context. For engineering leaders, that turns harness design into a cost question: how much of a token budget is spent gathering context versus producing useful output.

He was also blunt about current limits. Agents, he said, are not good at writing specs, and human judgment still matters in areas like review and verification. He added that technical debt is a real issue, describing a pattern inside Augment where agents are good at duplicating code. The company has responded with focused cleanup sprints using agents themselves, guided by explicit specs for reducing debt.

On cost, Perneti said teams may eventually split workloads between frontier models for the hardest tasks and open-source or smaller models for well-specified coding work. If that shift happens, he argued, token costs could fall significantly.

Ars notes that Anthropic and Augment are not necessarily measuring the same things. Cat Wu, Anthropic’s head of product for Claude Code, had previously said Anthropic did not see “a measurable improvement in performance” from adding some code-navigation tools such as LSPs. That is not the same as evaluating a broader context engine like Augment’s, which may help explain the differing conclusions.

What both sides appear to agree on is the larger direction: software development is changing quickly, more autonomous workflows are coming, and better models will not remove the need for engineering judgment. If anything, faster output may make bad decisions more expensive.