The race to build the best AI coding assistant is no longer just about the underlying model. It is increasingly about the software wrapper around it, a layer that determines what the model sees, what it can touch, and how it reasons through a codebase. Two prominent players, Anthropic and Augment Code, have landed on sharply different philosophies for this wrapper, and the gap between them is widening into one of the most consequential debates in developer tooling right now.
Anthropic, through its Claude Code product, has adopted what it calls a lean harness. Cat Wu, head of product for Claude Code, told Ars Technica earlier this summer that the company sees little value in pre-building structured context around a repository. The models are improving so rapidly, she argued, that locking in assumptions today risks obsolescence tomorrow. "Going by the evals, we don't see a measurable change," Wu said of more elaborate context systems. The team ships a thinner layer and lets developers bolt on what they need.
Augment Code sees the same exponential curve and draws the opposite conclusion. Vinay Perneti, the startup's VP of Engineering, believes that raw model intelligence is only half the equation. The other half is context, and gathering it on the fly burns tokens, time, and money.
Why Augment Code Pre-Indexes Every Repo
Perneti's team has spent roughly 18 months, starting before ChatGPT existed, researching how to retrieve the right slices of code from large repositories. The result is a system that pre-indexes a codebase using embeddings, a custom retrieval model, and a vector database, then serves up conceptually relevant code in sub-millisecond queries.
The payoff, Perneti claims, shows up most clearly in large private codebases. Public open-source repositories are a different story. Most frontier models have more or less memorized them through training, so a grep-based search is often enough to point the model in the right direction. Private repos are alien territory. The model has never seen them. Without semantic guidance, it wanders.
Augment Code published a head-to-head benchmark on Terminal-Bench using the same underlying model as Claude Code. The accuracy was comparable, but Augment Code completed tasks with 33 percent fewer tokens. Perneti attributes the savings to spending less time on exploratory context gathering.
The Measurement Problem
Wu and Perneti may not actually disagree on the facts so much as they are measuring different things. When Wu said Anthropic found no measurable gain from "a few LSPs available," she was referring to language server protocols, a narrower layer than Augment Code's full retrieval engine. Perneti was quick to point out that not all retrieval systems are equal, just as not all databases are equal.
"When somebody says, 'Hey, I tried Claude Code with a RAG implementation, but I'm not seeing benefits,' that's because the implementation and the context engine are very, very different," Perneti said. The two companies are optimizing for different outcomes: Anthropic for flexibility and future-proofing, Augment Code for token efficiency and speed in unfamiliar codebases.
What This Means for Engineering Leaders
The debate lands in the middle of a tension every engineering leader is feeling right now. Agents are getting good enough to write and edit code autonomously, but they are not free. Token budgets are real line items. Perneti framed the problem as a systems engineering challenge: how do you allocate energy between intelligence and context to get the best outcome for the least cost?
He also addressed the two loudest objections he hears from skeptical developers. The first is trust. Agents, he agreed, are poor at writing specifications. They need human judgment at the design and review stages. But once a clear spec exists, they execute well. The second is technical debt. Augment Code has observed internally that agents are "very good at duplicating code," which forced the company to run focused sprints to clean up the mess. The irony: you can write a spec for reducing tech debt, and the agent will execute that too.
The Open-Source Tipping Point
One variable that Perneti raised, and that Wu did not address, is cost pressure on frontier models. As models like Anthropic's Opus or Fable push deeper into agentic workflows, their price tags may become unsustainable as the default choice. Open-weight models, some now small enough to run locally, are closing the gap on routine coding tasks.
Perneti predicted a split workflow: frontier models reserved for the hardest problems, open-source or local models handling well-specified implementation steps. At that point, the cost equation shifts dramatically. Teams that have already built harnesses capable of routing tasks to the right model tier will be ahead.
The Real Winner May Be Judgment
Both Wu and Perneti expect frontier model capabilities to keep improving at or above the current exponential pace for at least the next year. Both agree this is rewiring how software gets built. Where they diverge is on whether the harness around the model should be minimal or richly opinionated.
A year from now, one approach may dominate, or the market may simply bifurcate by use case. Either way, the deeper implication is that faster, more autonomous production does not reduce the need for sharp engineering judgment. It amplifies it. A bad architectural decision executed by an agent at scale is a much bigger problem than a bad decision executed by a human at human speed. The teams that thrive will be the ones that treat agents as teammates, not replacements, and that invest in the systems that keep those teammates pointed in the right direction.
Watch for the next round of benchmarks comparing these harnesses on real private enterprise codebases. That is where the marketing claims will meet the metrics that actually matter.