When an AI assistant helps a programmer write code, it is essentially trying to solve a very difficult scavenger hunt. The tool needs to understand thousands of files at once to fix a bug or add a new feature. How it finds that information is currently a point of major debate among software developers.
Major AI companies and startups are building software known as a harness, which acts as a bridge between an AI model and a developer’s workspace. This bridge decides exactly what code the AI can see and how it should search for it. Some teams, like the one behind Anthropic’s Claude Code, prefer a lean harness—a simple, minimalist approach that avoids pre-sorting code. Others, like Augment Code, believe the only way to work on massive, internal company projects is to build a complex indexing system that catalogs the entire project before the AI even starts working.
The hunt for information
To understand why this matters, think of an AI model like a brilliant but forgetful intern. If you ask this intern to finish a task in a massive library, they need to know which books are relevant. One way to do this is for the intern to scan every single shelf every time they work, but that is slow and expensive. That is similar to a simple harness, which uses basic search methods to find keywords as needed.
Building an index is like creating a massive, searchable card catalog before the intern starts. This process uses a technique called embedding, which turns pieces of code into mathematical patterns that signify their meaning, rather than just their words. These patterns are stored in a specialized database, often called a vector database. When the intern needs information, they check the index first, allowing them to pull up exactly the right files in milliseconds. This is generally more reliable for private, proprietary software projects that the AI has never seen before.
This conflict boils down to a balance between quality and cost. Every time an AI reads a file, it consumes tokens, the units of text that companies pay to process. A complex index helps the AI find its target faster, which can save a significant amount of money and time. However, it also adds complexity to the software that keeps everything running. As AI assistants become more capable, the question for the industry isn't just how smart the next model will be—it is how we can ensure these models have the right information ready at their fingertips without burning through budgets.
Liked this one? The next lands at breakfast.
Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.
By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy