Guides · diagnosing the file-hunt tax · 6 min
Your agent isn't thinking slowly. It's looking slowly.
The lag you blame on the model is mostly disk. Before an agent can reason about your code, it has to find it: which file, which function, where it lives. On a real 4.47-million-file machine we measured where a session's time and tokens actually go, and more than half never reach your question. The bottleneck isn't intelligence. It's search.
Read the symptoms honestly
You're debugging the wrong layer.
Almost everything that feels like "the AI is slow today" is the same underlying event: the agent shelled out to a search tool and is waiting on the filesystem. It looks like thinking. It is mostly waiting. Here is the translation table.
| What you see | What you blame | What's actually happening |
|---|---|---|
| A long pause before it says anything | the model is thinking hard | it is walking the disk with find and grep |
| It gets slower the longer the session runs | the context is just getting long | every turn re-reads pages of earlier search output |
| You hit your usage limit by lunch | you simply used it a lot | over half your tokens went to directory listings |
| The bill outruns the work done | the model is expensive | you paid top-tier rates to read ls output |
| It gives up and asks you where the file is | it isn't smart enough | it spent its whole search budget guessing folders |
The right-hand column has one name: your agent is I/O-bound on file search, not compute-bound on reasoning. Reaching for a smarter or larger model doesn't fix an I/O bottleneck: it just pays more per second to wait on the same disk.
The centerpiece · where the tokens go
More than half the session never touched your question.
Every tool result the agent runs is appended to the conversation and re-read on the next turn. A find that returns thousands of candidate paths adds thousands of lines the model pays to read again, and again. Drawn as a token budget, one real session splits like this.
WITHOUT AN INDEX
WITH AN INDEX
Read it like a budget. The red slice is not one search: it is dozens of small ones (the notches), each a directory listing, a grep dump, or a re-read of candidate paths the model already saw. None of it is the answer you asked for. Give the file-hunt layer a real index and that slice all but disappears; the tokens you paid for go where you wanted them: reasoning about your actual problem.
One task · measured, not estimated
We watched Claude Code find one file.
That was the whole workflow (plan, search, narrow, repeat) walked call by call in how Claude Code finds files. Content search tells the same story: one representative query took 93.8 seconds on ripgrep (the fastest grep there is) where the index returned the identical result in 16 milliseconds. The agent didn't get smarter or dumber between those two numbers. Only the floor under it changed.
| The job | Without an index | With Interlinked |
|---|---|---|
| Claude Code finds one file | 6m 57s · 71 calls | 16 ms · 1 call |
| ripgrep runs a content query | 93.8 s | 16 ms |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. The Claude Code row is the measured agent run to open one file; the ripgrep row is a content query over the same corpus. The same disk walk is what VS Code and Windows Search do too: every tool here was built for one person with one project open, not an agent that questions the whole machine.
Metered or capped: same leak
You feel it two ways. Neither is the model's fault.
How the wasted tokens reach you depends only on how you pay. The waste itself is identical.
CLAUDE CODE · METERED
You pay per token. Every directory listing the model re-reads is billed at top-tier rates, so the file-hunt tax lands straight on the invoice, and stretches each task from seconds into minutes. The slowness and the bill are the same event, counted twice.
CLAUDE MAX · $200 SUBSCRIPTION
You pay a flat monthly rate against a usage limit. Tokens burned hunting for files are tokens off that limit, so you hit the cap sooner and wait, having spent a chunk of the allowance on ls output you'll forget one turn later.
Two pricing models, one root cause. Finding files is the layer every session stands on (the backbone of the LLM economy), and today it runs on tools built for humans, not agents. Whether you meet the tax as a bigger bill or a limit you hit before lunch, the fix is the same, and it is not a different model: stop making the agent walk the disk.
The fix · make the floor answer at memory speed
Hand the agent an answer, not a search.
Interlinked keeps a live index of every file on the machine (names and contents) and exposes it to any agent over MCP. No model, no embeddings, no cloud: just an index that answers offline and never leaves your machine. The agent asks once and gets the path or the matching lines back before it could have finished typing find /. No directory walk, no output flood, no re-reads. The token budget snaps back to reasoning.
SAME MACHINE · SAME MODEL · THE ONLY THING THAT CHANGED IS THE FLOOR UNDER IT.
Fix it today
Give your agent the floor back.
Install once. One signed installer auto-configures your AI clients (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed and the rest) over MCP. No JSON to edit.
Let it index. Filenames are searchable immediately; contents fill in behind, a light background service you'll forget is running.
Watch the shape change. The lookup that cost 71 tool calls becomes 1. The file-hunt slice of your token budget drops toward zero: the tokens go to the work.
Scope when you know where to look. Whole-machine when you don't; a folder when you do. Either way the answer is back before the model would have finished its first ls.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.