New arXiv paper cuts coding agents’ repository lookups by up to 87 percent

A new preprint posted to arXiv this week describes CodeNib, a data system built to give AI coding agents faster, cheaper access to the context of a software repository as it evolves. Rather than having an agent repeatedly re-scan files to understand a codebase, CodeNib maintains reusable lexical, dense and structural views tied to each commit, preserving those views through edits and mapping them back to precise source ranges so an agent can jump to relevant code without rebuilding its understanding from scratch. The authors report that graph and vector index updates ran 8.7 and 25.4 times faster at the median than rebuilding from nothing, verified against independent rebuilds across 100 repository snapshots. In a test of live coding requests, the system’s static navigation matched live server results with a median latency 4.7 times better across most of 1,000 sampled requests, and across five different language models, CodeNib’s context selection cut token consumption by 50 to 87 percent compared with conventional grep and read based approaches, without sacrificing how accurately the agent located the code it needed. The work is a narrow, technical contribution rather than a headline result, but it points at a real cost problem for the agentic coding tools now built into most major AI assistants: every extra token an agent spends rereading a repository is money and latency that a system like CodeNib is aimed squarely at cutting.