Agents · measured July 2026
Marc Benioff bet Salesforce on agents. Here is what agents do all day.
In 2024, Salesforce co-founder and CEO Marc Benioff (the man who spent 25 years selling the enterprise on software as a service) pointed his company at a single idea: AI agents. He launched Agentforce, started calling agents "digital labor," and set a public goal of a billion of them in use by the end of Salesforce's fiscal 2026. Take the bet seriously (Benioff has been early on platform shifts before) and one plain question follows: what does an agent actually do all day? So we measured it, on a real 4.47-million-file machine. The biggest line item was not reasoning. It was finding files: about 58% of the tokens in the session we clocked.
TL;DR · the 30-second version
- 01Marc Benioff pointed Salesforce at AI agents: the Agentforce platform, the "digital labor" framing, and a public target of a billion agents in use by the end of its fiscal 2026.
- 02Take the bet seriously and a plain question follows: what does an agent do all day? We measured it on a real 4.47-million-file machine (Ryzen 9 9950X3D, Windows 11).
- 03The biggest line item was not reasoning. It was finding files: about 58% of the tokens in one Claude Code session.
- 04Same task, index off then on: 6m 57s and 71 tool calls collapsed to 16ms and a single call.
- 05On a metered plan like Claude Max (up to $200/mo) that hunting is real money. A whole-machine index answering in microseconds hands the budget back to the work.
Take the bet seriously
He wagered the company on agents. So what is an agent?
In September 2024 Salesforce launched Agentforce, its platform for building AI agents, and Benioff reorganized the company's entire story around it. He began calling agents "digital labor," framed Salesforce's future as a provider of a digital workforce, and set a public target of a billion agents deployed by the end of its fiscal 2026. He bet on the cloud when the enterprise still shipped software on discs; the agent bet is the same shape.
An AI agent is a large language model placed in a loop: it reads a goal, calls tools to gather the files and facts it needs, acts, and checks the result, repeating until the job is done. Strip away the branding and most of an agent's day is that one middle step: gathering what it needs. That step is retrieval, and it is the same reason a solo founder can now run a swarm of agents, the argument in the one-person company runs on retrieval. The swarm still has to find things.
So we pointed a stopwatch at the least glamorous step and measured where the time and the tokens actually went. The vision and the measurement did not match.
The vision vs the measured day
The pitch was pure work. The stopwatch disagreed.
The top bar is the day Agentforce sells: an agent reasoning, planning, acting, delivering. Labor, digitized. The middle bar is the day we clocked on a real machine. Same agent, same kind of task; most of the token budget never reached the problem. It went to locating files and reading them back.
Same task, same machine, twice. Without a whole-machine index the agent did directory archaeology (list a folder, read the bytes, guess, repeat) for nearly seven minutes and 71 tool calls to reach a single file that was already sitting on the disk. Seven minutes is long enough to make a coffee and drink half of it; 16 milliseconds is faster than you can blink: one call instead of 71, roughly 7,200,000× less work for the same file. This is the argument in full in the backbone of the LLM economy, and the economics of it in software ate the world, now agents eat your tokens.
Same machine · same drive · same queries
The retrieval floor, priced in seconds.
Every row is the same errand: fetch one file off a 4.47-million-file drive. Windows Search takes 67 seconds, long enough to refill your coffee; the indexed lookup takes 139 microseconds, and the answer is on screen before you register the wait.
| The searcher | What was measured | Time | vs Interlinked |
|---|---|---|---|
| Interlinked | one indexed lookup, 4.47M files | 85 μs | baseline |
| VS Code Ctrl+P | find a file by name, same drive | 35 s | 412,000× |
| Windows Search | 20 filename queries, median | 67 s | ~480,000× |
| ripgrep (what your agent runs) | content search, same corpus | 93.8 s | 58,625× |
| Claude Code (no index) | the full file hunt, 71 tool calls | 6 m 57 s | ~7,200,000× |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Filename lookups are median-of-20 (139 μs); the 85 μs figure is a single-file lookup, best case 2 μs. The ripgrep row is a content query over the same corpus: 58,625× is the average across the query set, and one query from it ran 93.8 s → 16 ms. The Claude Code row is the measured agent hunt (6 m 57 s, 71 tool calls) against a single indexed lookup; with the index the same task was one call, 16 ms. Full per-tool write-ups: the 412,000× VS Code comparison and the ripgrep content benchmark.
The economics nobody prices in
You can't out-model a slow floor.
A bigger model reasons better per token. It does not find files faster. It shells out to the same primitives underneath: list the directory, read the bytes, guess again. The agent Benioff's customers actually reach for to write code (Claude Code on an Anthropic Claude Max plan, up to $200 a month) is metered in tokens, and the meter runs the entire time the agent is looking. Most of what an agent does is look.
So if about 58% of a session goes to finding files, more than half the plan buys directory listings the model reads once and forgets. A faster model would only forget them faster. An always-fresh index that answers machine-wide in microseconds turns that floor into bedrock. The token math is in where your tokens go. Put plainly: agents don't need bigger brains as much as they need to stop looking for their keys.
The other half of the bedrock
Speed finds the file. Memory knows what it's for.
The same index that answers in microseconds also carries a knowledge vault it builds by itself: relationships derived from your own history, plus notes any agent can leave on any file that every other agent then reads. It is keyed to the file's OS identity, so it survives renames, spans every repo, and belongs to no vendor. That is the digital workforce's missing memory: not the model's recall of your chat, but the machine's recall of itself. It stays about a millisecond behind your last save, sits around 44 MB at rest (less than a browser tab) and takes under 1% of the drive.
EVERY EDGE IS A RECEIPT: A COUNT OR A HASH YOU CAN CHECK, NEVER A MODEL'S GUESS.
FAQ
Questions people actually ask.
What is Agentforce?
Agentforce is Salesforce's platform for building and deploying AI agents: software that takes actions across a company's systems, not just answers questions. Salesforce launched it in 2024 and set a public target of roughly a billion agents in use by the end of its fiscal 2026, with Benioff reframing the category as "digital labor."
What does an AI agent actually spend its time on?
On the session we measured, most of it was retrieval: locating the right files and reading them back. About 58% of the tokens went to finding context, not reasoning about the problem. A faster model does not fix that; it shells out to the same slow file search underneath. We break the split down in where your tokens go.
How much does file-hunting cost on a plan like Claude Max?
Claude Max is metered in tokens and tops out around $200 a month. If roughly 58% of a session's tokens go to finding files, more than half of that bill buys directory listings the model reads once and forgets. The fix is a faster retrieval floor, not a bigger model.
How is Interlinked different from my editor's search or a vector database?
Editors index the one project you opened; a vector database stores embeddings for fuzzy semantic recall. Interlinked is a machine-wide index (every repo, every folder, including the git-ignored files a project search misses) that answers in microseconds and stays about a millisecond behind your last save. It sits underneath both, not in place of them. See what an MCP server is for how agents plug into it.
Is any AI running on my machine?
No. There is no model in the box: just an index and a lookup, offline and private. That is why it can answer in microseconds, and why the relationships it builds are receipts (counts and hashes you can check) rather than a model's guess.
Do this today
Give your agent the bedrock.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing.
Let it index. Filename search works immediately; content fills in behind it. Freshness stays about a millisecond behind your saves, under 30 ms at worst.
Point it at everything. Whole-machine when you don't know where something lives; scoped to a folder when you do: scoped queries return in microseconds.
Watch the meter drop. The tokens that were spent finding files stop being spent finding files. That budget goes back to the problem you were actually solving.
Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo. Download for Windows.