Can an AI coding agent be faster, cheaper, and more accurate by reading less code?

I built Faber around this question.

Most coding tasks only depend on a small part of a repository, but an AI coding agent can end up reading a lot of unrelated code. That means more tokens, higher API cost, more latency, and potentially more irrelevant context for the model.

Faber approaches this differently. It builds a code graph and uses relationships between files, symbols, callers, and dependencies to narrow down what matters before loading context.

The idea is:

• Less irrelevant code → fewer input tokens
• Smaller context → lower API cost
• Targeted search → faster repository navigation
• More relevant context → potentially better answers
• Prompt caching → avoid repeatedly paying for the same context

Faber also exposes /usage, so you can actually see tokens, caching, cost, and estimated savings instead of treating model usage as a black box.

I’m curious what others think: as context windows keep getting larger, should coding agents keep reading more code, or get better at deciding what not to read?

GitHub: https://github.com/JibanKumar-cloud/faber

More on the design and code-graph approach: https://medium.com/@jshial25/why-should-an-ai-coding-agent-read-hundreds-of-files-to-answer-one-question-d6369d29dfa5