Opinion · agents · measured July 2026
The AI coding bottleneck isn't the model.
The models got fast and got smart. Watch a coding agent work, though, and the slowest thing it does all session isn't thinking. It's finding the file. We put a stopwatch on it: one lookup that took Claude Code 6 minutes 57 seconds, and roughly 58% of a session's tokens, spent not on code but on hunting for it. The ceiling keeps rising. The bottleneck is the floor.
TL;DR
Key takeaways. Same model, faster floor.
The bottleneck isn't the model. It's the agent finding the file. One lookup cost Claude Code 6m 57s across 71 tool calls; a warm index answers it in 16 ms with a single call, about 7,200,000× faster.
Roughly 58% of a coding session's tokens go to hunting for files instead of writing code. Give the agent a machine-wide index and that collapses to ~0%.
A single filename lookup returns in 85 μs (139 μs median across 20 queries), 412,000× faster than VS Code's Ctrl+P and ~480,000× faster than Windows Search's 67 s.
Content search runs in 7 to 16 ms where ripgrep's heavy query takes 93.8 s: the proof query collapses 93.8 s → 16 ms, 58,625× on the same corpus.
The index stays fresh ~1 ms after a save and idles at ~44 MB. Everything local is free forever, no card. Hosting starts at $5.99/mo.
Where the session actually goes
The model's job is the small part. The rest is looking.
A coding session's token budget is not mostly reasoning. On a real 4.47-million-file machine, the majority of what an agent spends goes into locating and reading files before it can even start, reading walls of directory output it will forget one turn later. Draw the budget to scale and the shape of the problem is obvious: the expensive layer isn't the model, it's everything that has to happen before the model can help.
Two famous tools · one stopwatch
Claude Code and VS Code, on the same machine.
Neither of these is a bad tool. The agent is smart; the editor is fast at the job it was built for. But the moment either one has to answer "where does this live on the whole machine?" it falls to the same primitive (walk the disk, scan bytes, hope) and the wall-clock explodes. A faster model on that floor is a faster car in the same traffic.
The argument
Fix the floor, not the ceiling.
Almost all of the effort in AI coding goes into the ceiling: bigger models, longer context windows, cleverer prompting, better harnesses. All real, all useful. But the ceiling was never what capped a run. The floor is: the physical act of the agent locating and reading the right files. That layer hasn't moved. It's the same directory-walking and content-scanning a laptop used in 2006, and every model you buy still stands on it. Here is the same lookup, everywhere it happens today.
| The searcher | The task | Time | vs Interlinked |
|---|---|---|---|
| Interlinked | one indexed lookup | 85 μs | baseline |
| Claude Code (agent, no index) | find one file, end to end | 6 m 57 s | ~7,200,000× |
| VS Code (Ctrl+P) | open a file you can name | 35 s | 412,000× |
| ripgrep | one heavy content query | 93.8 s | → 16 ms |
| Windows Search | median of 20 queries | 67 s | ~480,000× |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Filename lookups are timed against Interlinked's 85 μs single-file lookup (139 μs median across 20 queries); content and agent runs collapse to 16 ms. The ripgrep row is the proof query, 93.8 s → 16 ms; across the 5-query content set it is 58,625× on the same corpus. Full per-tool methodology in the 412,000× write-up, the ripgrep comparison, and the Claude Code agent run.
The floor, rebuilt
A save is searchable in about a millisecond.
The fix isn't a better model. It's a faster floor. One background index of the whole machine (every repo, plus the roughly 90% of the disk that was never in any repo: configs, documents, downloads, the PDF with the answer) kept current as you work. Edit a file and it's searchable in about a millisecond, under 30 milliseconds in the worst case. The agent stops walking directories because, for the first time, there's something warm to ask.
And it costs almost nothing to keep around: about 44 MB of memory at rest (less than a browser tab) with the index on disk staying under 1% of the drive (~24 GB on a 4 TB machine). One signed installer wires it into 19 AI clients at once. The model you already pay for gets to spend its whole budget on the work: make the layer every agent stands on answer at memory speed.
Raising the floor, part two
Speed finds the file. Memory says what it's for.
Finding fast is only half the floor. The same index also keeps a knowledge layer it builds by itself (which files change together, which repos are one shipping unit, which project went quiet) plus notes any agent can leave on any file that the next agent reads. No model wrote a word of it; every relationship is a count or a hash you can check. It's keyed to each file's identity on disk, so it survives renames and spans every repo. Not a replacement for your model's memory or a vector database: the layer underneath both.
THE FLOOR IS SPEED PLUS MEMORY, AND BOTH SHIP IN ONE INSTALL.
Common questions
Questions, answered.
Isn't a smarter or faster model the real fix?
No. The model was never the slow part. The same lookup a warm index answers in 16 ms cost Claude Code 6m 57s across 71 tool calls, and about 58% of a session's tokens go to finding files rather than writing code. A faster model just drives faster in the same traffic.
How is this different from ripgrep, VS Code search, or Windows Search?
Those walk the disk or scan bytes the moment you ask; Interlinked answers from a prebuilt, always-fresh index. A filename lookup returns in 85 μs (412,000× faster than VS Code's Ctrl+P and ~480,000× faster than Windows Search) and a heavy content query that takes ripgrep 93.8 s collapses to 16 ms.
Does running a whole-machine index slow my computer down?
No. At rest the engine idles at about 44 MB (less than a browser tab) and a save is re-indexed in about 1 ms, so results stay current. It's a background index, not a scan that fires every time you search.
Is there an LLM or any AI model in the box?
None. Every result is an index lookup plus a path/id match, and every relationship it tracks is a count or a hash you can verify. Nothing leaves the machine, and no answer is a model's guess.
What does it cost, and which tools does it cover?
Everything (including agent access, the MCP layer that wires into 19 AI clients like Claude Code, Cursor and Copilot) is free forever, on all your devices, no card. Hosting starts at $5.99/mo.
Do this today
Give the model a faster floor.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing, no keys.
Let it index. Filename search works immediately; content search fills in behind it. At rest the whole engine sits around 44 MB, less than a browser tab.
Scope when you can. Whole-machine when you don't know where something lives; scoped to a folder when you do. Either way the answer comes back in the microsecond-to-millisecond range.
Stop paying the model to grep. The tokens your agent burned hunting for files go back to reasoning and writing code. The bottleneck moves off the floor, where it belonged all along.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.