AI Burn Clock Tracks the Cost of AI Agents Reading Too Much

Summary

AI Burn Clock is a project about the token and financial cost of AI agents that answer questions by opening whole files instead of locating the relevant passage first. In three questions on one production Next.js codebase, the site reports that whole-file reading loaded 16 to 47 times more bytes than an index-first workflow: 103 KB fell to 6 KB, 244 KB to 6 KB, and 242 KB to 5 KB. It also cites a published coding case study reporting 2.7 times fewer output tokens with retrieval-first, while using a more conservative 63% avoidable-reading factor for its estimates. Applying its reference factors to reported spending, the site models avoidable reading across worldwide AI spending, U.S. federal AI obligations, JPMorgan Chase’s reported AI budget, and one vendor’s reported revenue run rate; it labels these figures scenarios rather than direct measurements. Its proposed remedy is XERJ, an Apache-2.0 open-source Rust search engine that indexes a folder locally, returns files, lines, and symbols, and can work with MCP and an agent-memory plugin. The page says the index runs on the agent’s machine and is intended to answer “where is it?” before the agent reads the pointed-to passage.