Field notes · agents · measured July 2026
Skynet went self-aware. Your agent still can't find its own config file.
Science fiction kept promising the same machine: one that wakes up and instantly knows everything. The machines did wake up; they reason like prodigies. Then you ask a top agent to find a single file on a real 4.47-million-file drive, and it spends 6 minutes 57 seconds and 71 tool calls groping in the dark. Intelligence was never the bottleneck. Retrieval is.
TL;DR · the 20-second version
- 01Sci-fi sold us self-aware superintelligence. The real 2026 agent is brilliant at reasoning and blind to your filesystem.
- 02Asked to find one file on a 4.47M-file machine, a top agent (Claude Code) spent 6m57s and 71 tool calls, and burned ~58% of its session tokens hunting, not thinking.
- 03Intelligence was never the bottleneck. Retrieval is. A smarter model just reads the wrong directory listing faster.
- 04An always-fresh, whole-machine index closes the gap: the same task in 16 ms with one call (~7,200,000×), a single filename in 85 μs.
- 05It also hands over a map: 8,455 git events read into 26,958 relationships in 1.7 s, so the next agent starts oriented.
The doorway, drawn honestly
The fiction imagined omniscience. The reality forgot where it put things.
In the Terminator films, Skynet is a fictional defense network that becomes self-aware, decides humanity is a threat, and coordinates a global war in an instant. It is our culture's shorthand for the same fantasy: a machine that wakes up and, in one flash, comprehends and acts on everything.
The machines that actually woke up in our timeline are astonishing reasoners. But they ship blind to your filesystem. Ask one where a file lives and it falls back to the tools a 2006 laptop used (walk directories, run grep, read walls of output, hope) because it has no map of your machine. Skynet is the doorway here, a metaphor and a joke, not a claim that any agent is it, and not an endorsement. The point is the gap between the fantasy and the mundane failure below, and the measured layer that closes it.
One real task, on the clock
We watched a genius look for one file.
That is ~7,200,000×. Put plainly, nearly seven minutes of an agent typing ls, reading output, and forgetting it a turn later, versus 16 milliseconds: shorter than the blink you would use to look away. The model was never the slow part. It reasoned brilliantly about the wrong directory listing, 71 times. Tokens spent groping are tokens not spent solving. On a $200/month agent plan, more than half the bill can go to hunting. The full teardown lives in the Claude Code experiment and why one lookup becomes 71 calls.
Same machine · same drive · same file
One file, found every slow way there is.
| The searcher | The task | Time | vs Interlinked |
|---|---|---|---|
| Interlinked | one indexed lookup | 85 μs · 16 ms | baseline |
| Claude Code (no index) | find one file (the agent hunt) | 6 m 57 s | ~7,200,000× |
| ripgrep (the agent's fallback) | same content query, same drive | 93.8 s | 58,625× |
| VS Code Ctrl+P | same file, same drive | 35 s | 412,000× |
| Windows Search | median of 20 queries | 67 s | ~480,000× |
Multipliers are canon pairings: the agent workflow and the ripgrep content query compare to Interlinked's 16 ms; single-file lookups (VS Code, Windows Search) compare to its 85 μs / 139 μs. Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Full per-tool method in the ripgrep comparison and the token-cost breakdown.
Why a bigger brain didn't fix it
Retrieval is a different axis from intelligence.
Retrieval is the part of an agent's job that finds the right file and the right lines before the model can reason about them. It does not get faster when the model gets smarter. It gets faster when there is an index. VS Code knows the folder you opened. Cursor knows the repo you are in. Your machine has a dozen repos that ship together, plus the 90% of the disk that was never in git: configs, documents, downloads, the PDF with the answer. An agent's questions don't respect folder boundaries; its tools do.
So the smartest model on earth shells out to the same floor as everything else: ripgrep, walking the disk on every single query, because nothing warmer exists machine-wide. An always-fresh index turns that floor into bedrock. That is the whole idea, and it is the umbrella thesis of the backbone of the LLM economy: make the layer every agent stands on answer at memory speed.
The measured layer that closes the gap
Give the genius a map. Watch 6m57s become 16ms.
Interlinked is a pre-built, always-fresh index of every file on the machine: names and contents, whole-machine, every repo, including the git-ignored files a working-directory grep misses. No model, no embeddings, no cloud: the answer is an indexed lookup, so it survives offline and never leaves the machine. The same index carries a knowledge vault (relationships it derives by itself, keyed to each file's OS identity so they survive renames) plus notes any agent can leave for the next.
NO LLM IN THE BOX · EVERY EDGE IS RECEIPTS: COUNTS AND HASHES YOU CAN CHECK, NEVER A MODEL'S GUESS.
Straight answers
Questions the skeptic asks.
Can't a smarter model just find files faster?
No. Intelligence and retrieval are different axes. A better model reasons brilliantly about the wrong directory listing, faster; it still has no map of your machine. On a 4.47M-file drive that means 71 tool calls and 6m57s to reach one file. The fix is an index, not IQ.
Is this a knock on Claude Code, or on Terminator?
Neither. Claude Code is the agent we make faster (the hero of the story, not the villain). Skynet is a playful metaphor for the sci-fi promise of a machine that instantly knows everything. The joke is the gap between that fantasy and a real agent groping for a config file. It is a doorway, not a technical claim and not an endorsement.
How fast is the fix, really?
A single filename lookup lands in 85 microseconds (best case 2), with a 139-microsecond median across 20 queries on 4.47M files. Content search returns in 16 milliseconds where ripgrep took 93.8 seconds: 58,625× on the same query. The full agent task collapses from 6m57s and 71 calls to 16 milliseconds and one call.
Does the index go stale between saves?
No. Save a file and it is searchable in about a millisecond, under 30 milliseconds in the worst case. The engine watches the filesystem for changes and keeps up live. At rest the whole thing sits around 44 MB, less than a browser tab.
What do I install, and what does it cost?
One signed Windows installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Cline and the rest. No JSON editing. Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo.
Do this today
Give your agent the map.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Cline and the rest. No JSON editing.
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.
Ask the hard question. 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.
Let the vault grow. Day one it knows your git history: 26,958 relationships from your own repos. Every week after, it knows your machine better, automatically.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows and let the next session start with the map already drawn.