Beyond grep: The case for a context-rich AI coding harness

There's a quiet argument happening inside the AI coding world, and it's not about which model writes better code. It's about what happens before the model even starts typing: how does an AI agent figure out what's actually in your codebase?

Anthropic's Claude Code team, represented by product lead Cat Wu, favors what they call a lean harness — minimal scaffolding, mostly grep-based search, and a bet that the underlying models are improving so fast that heavy infrastructure will just become dead weight. Augment Code, a startup that's been building context systems since 2022 (before ChatGPT even existed, notably), takes the opposite approach. Its VP of Engineering, Vinay Perneti, told Ars Technica that Augment pre-indexes an entire repository using embeddings, a dedicated retrieval model, and a vector database tuned for sub-millisecond lookups. The idea: don't make the agent rediscover your codebase from scratch every single task.

Perneti's strongest argument centers on private repos. Public benchmarks tend to run on open-source code that large models have effectively memorized during training, so of course a grep-based agent looks fast and competent there — it already half-knows the answer. Private, proprietary codebases are a different story entirely, since no model has ever seen them before. Augment claims its semantic retrieval closes that gap, and points to a Terminal-Bench test where it matched Claude Code's accuracy while using 33 percent fewer tokens.

So why does Anthropic's own testing show no measurable gain from adding code-navigation tools? Perneti's answer is basically: not all retrieval systems are created equal, and Wu was likely testing something narrower, like a language server protocol, rather than a full context engine with years of research behind the embedding and retrieval models. It's a fair point, though also a convenient one for a company whose whole product is that engine.

What's more interesting is where Perneti takes the cost conversation. He argues that as open-weight models keep closing the gap with frontier ones, well-specified coding tasks may increasingly get routed to cheaper, even locally-run models, while the priciest frontier systems get reserved for the genuinely hard problems. If that shift happens, the harness — the layer deciding what a model sees and how it works — becomes the thing that actually determines your token bill, not just the model you picked.

Both camps agree models will keep improving fast, and both agree software development is already changing shape underneath developers' feet. The dispute is really just about timing and trust: build the scaffolding now and hope it pays off, or wait and let the next generation of models make the scaffolding unnecessary. Neither side has proven itself wrong yet.