The greps searched inside files already located; the reads and edits were the work. The only calls that reached across the whole machine to find a file were the five in blue.

The raw calls

Every whole-machine lookup: real numbers, tidier layout.

Reformatted for reading: the queries, filters, timings, and results are exactly what the tool returned. Note the scope and extension filters, the sub-millisecond timings, and the graph riding along on each hit (sibling repos and shared commit-words), so the agent knows which result matters before it opens anything.

iFileNames · 5 calls · session 4ea4644aall under 3.2 ms

1● interlinked · iFileNames (MCP)

2    query "style"   scope 1Interlinked   ext [md, mdx, tsx]   limit 30

3  → Found 1 result in 1.52 ms   (1,516,300 ns)

4    …/CellsInterlinked/src/app/research/BLUEPRINT_STYLE.md   18.8 KB   2026-07-03

5      keywords: hero · accent · toolbox · white · pane · tweak

6      <-> same-work  · the engine repo  · shares 583 rare commit-words

7      <-> same-owner · 4Angles     · shares 3 rare commit-words

8    // the canonical style doc, in a sub-path a cwd grep never reaches

9

10● interlinked · iFileNames (MCP)

11    query "popup"   scope 1Interlinked   limit 30

12  → Found 2 results in 3.20 ms   (3,200,400 ns)

13    …/components/desktop/diff-popup.tsx   16.4 KB

14    …/components/stats-popup.tsx           9.9 KB

15

16● interlinked · iFileNames (MCP)

17    query ""   scope …/src/components/homepage   limit 40

18  → 16 results   // the whole folder, ranked by modified (instant)

19

20● interlinked · iFileNames (MCP)

21    query "homepage"   scope 1Interlinked   modifiedAfter 6d

22  → 0 results   // a fast "no" is still an answer

23

24● interlinked · iFileNames (MCP)

25    query ""   scope .claude/projects   ext [jsonl]   modifiedAfter 5d

26  → 0 results   // the index skips .claude by design

The find that mattered

One call. Not seventy-one.

The file the agent needed lived in a different sub-path than the one it was working in: the kind of file a folder-scoped grep silently misses. The index returned it in 1.52 ms, with the graph attached. Our own benchmark for one un-indexed file lookup: 71 tool calls, 6 min 57 sec.

1vs71calls to
find one file

tool_result · from the log, reformatted1.52 ms

1● interlinked · iFileNames (MCP)(query: "style",

2    scope: "C:\\_Snow\\Coding\\1Interlinked",

3    extensions: ["md","mdx","tsx"], limit: 30)

4

5Found 1 result in 1.52ms (1516300ns):

6C:\\_Snow\\...\\CellsInterlinked\\src\\app\\research\\BLUEPRINT_STYLE.md  18.8 KB  2026-07-03 17:19

7  folder CellsInterlinked: keywords: hero, accent, toolbox, white, pane, tweak · active now · pushes to vmaly

8  <-> same-owner · C:\\_Snow\\Coding\\4AnglesGit\\4Angles · shares 3 rare commit-words · same remote account

9  <-> same-work  · C:\\_Snow\\Coding\\1Omega... · shares 8 rare commit-words

The momentum

Finding was a rounding error. Then it just built.

Five calls located everything the session touched. The other 194 were reads, edits, writes, and checks (the actual work). On our homepage we show the reverse: an un-indexed agent burns ~58% of a session just finding files. Here it was 2.5%.

194 calls building: reads · edits · writes · checks

↑ finding · 2.5%building · 97.5% →

Time to find one file, drawn to scale

Un-indexed hunt6m 57s

Interlinked< 15 ms

The blue bar is real. At this scale it is 0.004% as long. Same job, same machine.

What it saved

Five finds. Under 15 milliseconds, total.

At the industry rate for one un-indexed lookup (71 calls apiece), those five finds would have cost on the order of ~350 tool calls and half an hour of grepping. The index turned that into a rounding error, and left the whole call budget for building.

≈ 11 ms

total time across all 5 whole-machine finds

1 : 71

calls per file, indexed vs the benchmark hunt

~350

tool calls it never spent searching

The search is not the feature. The search is what makes every other call worth spending.

Give an agent instant retrieval, and it stops hunting.

The session happened to be tidying a design system. The topic is beside the point. What the log shows is the shape of every agent session once search is free: five instant lookups, then a long, uninterrupted run of real work. That is the layer under everything your AI does on a real machine.

Put it on your machine, free forever

Dissected from one real session · 199 tool calls · July 23, 2026← All research