Interlinked.

Guides · ranked & measured July 2026

7 things quietly slowing down your AI coding agent, and the one that dwarfs the rest.

Smaller prompts, the right model, tighter scope: the usual advice all helps at the margin. But on a real 4.47-million-file machine we ranked where an agent's time and tokens actually go, and one sink is bigger than the other six put together: finding files. Here they are, worst first, each with a fix.

The centerpiece · one session, ranked

One bar is more than half the session. The other six split the rest.

Read it like a budget. Each faint track is 100% of one agent session; each bar is what a sink actually costs. The measured split is ~58% on finding files, ~42% on everything else, so the single red bar outweighs the six gray ones combined. That is the whole point of the list: fix the top bar and you have done more than fixing all six below it.

WHERE ONE AGENT SESSION GOESBAR = SHARE OF THE SESSION01File huntinggrep · find · re-walking the disk~58%HALF THE SESSIONTHE OTHER SIX: 42% BETWEEN THEMranked by impact, not to scale02Re-reading outputevery listing re-read each turn03Standing contextfat prompt files, pasted trees04Serial tool callsreads run one at a time05Over-thinking editsmax effort on a rename06Vague scopewhole-machine when a folder would do07Dead-end retriesguess a path, miss, repeat~58% of session tokens on finding files, ~42% on everything else · Ryzen 9 9950X3D · Windows 11 · 4.47M files

Only the top split is a measured share (about 58% of session tokens spent hunting, on the benchmark machine); the ordering of the six below is by impact, not drawn to scale. The numbers behind the red bar (and the two anchors most people know, Claude Code and ripgrep) come next.

The list · worst first, each with a fix

Seven sinks. Six are habits. One is infrastructure.

None of these is a trick or a secret flag. They are the honest set of things that stretch a coding session, ordered by how much they move the clock. Numbers 02 through 07 you can improve this afternoon. Number 01 is a different kind of fix, and it is most of the problem.

01

File hunting~58% OF THE SESSION

The agent has no map of your machine, so it shells out to grep, find and ripgrep and re-walks the disk on every query. It is most of the wall-clock and about 58% of the tokens in a session.

FIX →Hand it a prebuilt, always-fresh index. A lookup becomes one call that returns the path or the matching lines, not a directory walk. The rest of this page measures it.

02

Re-reading stale output

Every tool result (a four-thousand-line directory listing, a grep dump) is appended to the conversation and re-read on every following turn. The session gets slower the longer it runs.

FIX →Don't create the wall. An indexed lookup returns an answer, not thousands of candidates; and compact or restart before context fills with listings the model will only re-read.

03

Bloated standing context

A fat instructions file, pre-pasted file trees and a giant system prompt are re-sent every single turn. You pay to re-read them each time, before the model even reaches your question.

FIX →Keep the standing prompt to what the model actually needs, and let it pull specifics on demand instead of pasting them in up front.

04

Serial tool calls

Independent reads and edits fired one at a time are dead wall-clock: each round-trip waits on the last for no reason.

FIX →Batch independent operations into one turn. Parallel tool calls collapse the waiting to a single round-trip.

05

Over-thinking small edits

Maximum reasoning effort on a one-line rename is latency you paid for and didn't need.

FIX →Match the model and effort to the task: deep reasoning for real design work, light and fast for routine edits.

06

Vague scope

“Find the auth guard” makes the agent search the whole machine; “the auth guard in packages/api” does not. A vague ask widens every search behind it.

FIX →Say where, not just what. A scoped path is the cheapest hint you can give, and against an index, a scoped query returns in microseconds.

07

Dead-end retries

With no map, the agent guesses a path, misses, and searches again. Each miss compounds the file-hunt tax at the top of this list.

FIX →An index doesn't guess. It has the file or it doesn't, first try, so there are no blind retries to pay for.

Number one · measured, not estimated

Two names you already know sit at the top of the list.

When a coding agent needs to search, it reaches for a scanner (most often ripgrep) and that scanner has no index, so it re-reads the bytes on every query. When the agent is Claude Code, that floor becomes the whole session. Both were measured on the same machine, against the same drive, doing the same work.

ANCHOR 1 · RIPGREP

93.8 s → 16 ms

the same content query: scanner vs index

58,625× faster across the five-query set (5.6 ms average). ripgrep is excellent software: it just re-walks the tree and re-reads the bytes on every single query, because it has no index to ask.

ANCHOR 2 · CLAUDE CODE

6m 57s → 16 ms

71 tool calls → 1 · ~58% of tokens → ~0%

The full find-a-file workflow, end to end: about 7,200,000× faster. Even a single isolated lookup took the agent 193.5 s where the index returned it in 85 µs, a 2,276,000× gap on one file.

Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Filename lookups are 85 µs single / 139 µs median across 20 queries; the ripgrep row is a content query (93.8 s on the heaviest of the set → 16 ms indexed). Full per-tool methodology in the ripgrep comparison and the Claude Code run.

Fix one, fix four

Four of the seven are the same problem in four disguises.

Look at the fixes again. Three of the seven are prompt discipline you own: trim the standing context (03), batch your calls (04), right-size the effort (05). The other four collapse into one root cause. Remove the file-hunt floor and 01 disappears, 02 has no output wall to re-read, 06 becomes a microsecond scoped lookup, and 07 stops guessing. One install, four of the seven.

ONE PREBUILT INDEXalways-fresh · one call01File huntingone indexed call02Re-reading outputno wall to re-read06Vague scopemicroseconds, scoped07Dead-end retriesno blind guessesThe other three (standing context, serial calls, over-thinking) are prompt habits, not infrastructure.

The fix for the top of the list

Give the agent an answer, not a search.

Interlinked keeps an always-fresh index of every file on the machine (names and contents, every repo, the git-ignored files a cwd grep never sees, documents included) and hands it to any agent over one call. No directory walk, no output flood, no blind retries. A save is searchable in about a millisecond, so the index never lies about a file you just wrote. One signed install auto-configures 19 AI clients (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed and the rest) with no JSON editing.

139 µs
median filename lookup across 20 queries on 4.47M files, median, not best case
1 to 10 ms
a content search across the whole machine, 5.6 ms average on the test set
~1 ms
from saving a file to it being searchable; under 30 ms worst case
~44 MB
idle footprint, less than a browser tab; the index stays under 1% of the drive

The same index carries a knowledge vault that rides along on the searches the agent already runs, with no model in the loop. On a fresh machine it found 12 of 12 repos and replayed 8,455 git events in 1.7 s into 26,958 relationship pairs, so the agent knows which files change together before it asks. Speed finds the file; the vault knows why it matters.

SAME MACHINE · SAME DRIVE · 412,000× FASTER THAN VS CODE, ~480,000× THAN WINDOWS SEARCH, 58,625× THAN RIPGREP.

Do this today

Clear the top of the list first.

1

Fix number one with one install. A signed installer wires up 19 AI clients (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed and the rest) over one call. That removes 01 and shrinks 02, 06 and 07.

2

Let it index. Filename search works immediately; content fills in behind it. At rest the whole engine sits around 44 MB (less than a browser tab) and the index stays under 1% of the drive.

3

Then spend ten minutes on the habits. Trim the standing context, batch independent tool calls, and stop maxing effort on trivial edits. Those are 03, 04 and 05, free once you look for them.

4

Say where when you know. Whole-machine when you don't know where something lives; a scoped path when you do, scoped queries return in microseconds.

Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.

Measured July 2026 · Ryzen 9 9950X3D · 4.47M files · Windows 11← All research