AI coding tools debate how much context to gather before coding

In short: AI coding tools are splitting between “lean” assistants that search code as they go and “context-rich” systems that prepare a map of your code in advance.

What's going on

AI coding tools are not only about the AI model itself. They also depend on the “harness,” which is the software wrapper that decides what the AI can see and do in your project (like a set of rules and tools around the AI).

Anthropic’s Claude Code team has argued for a “lean harness.” The idea is to keep built-in tools minimal, because AI models are improving quickly and extra structure may not show clear gains in tests. In this approach, the tool often searches the code when needed, similar to using a basic search function to find the right files.

Augment Code is betting on the opposite approach. Its engineering leader Vinay Perneti says Augment “pre-indexes” a company’s codebase. This means it builds a searchable map ahead of time using “embeddings” (a way to turn text into numbers so a computer can find similar ideas) and a “vector database” (a storage system built for those number-based lookups). Perneti says this helps most with large private codebases, where the AI has not seen the code before.

What to watch

A key question is cost and speed. Perneti claims Augment can finish similar tasks with fewer “tokens” (the bits of text an AI reads and writes, which often drive the bill), while Anthropic says it has not seen measurable improvements from some added code-navigation tools. Another open question is whether teams will shift more routine work to cheaper, smaller, or locally run models, while saving expensive top-tier models for the hardest problems.

Source: Arstechnica