---
title: Pay Attention
subtitle: attention as currency
status: draft
type: paper
created: 2026-08-19T00:00:00.000Z
paper: iii of iii
share_slug: workbench-attention-as-currency
tags:
  - thesis
  - workbench
  - paper
  - attention-as-currency
  - token-economics
  - model-choice
  - lock-in
  - holograph
  - landing-page
  - paper-iii
  - cost-model
  - retrieval-efficiency
  - orchestration
  - frontier
  - resource-scarcity
  - ai-economics
  - user-agency
  - architectural-design
  - cognitive-load
related:
  - "[[workbench-papers-editorial-plan]]"
  - "[[workbench-coordination]]"
  - "[[editorial-voice-style-guide]]"
  - "[[MDB-22.5-positioning-thesis]]"
description: Third paper exploring attention as scarce resource in AI work,
  examining cost models, token economics, and user choice in workbench
  architecture.
mdb:
  index_tier: 2
entity_id: MDB-6.5.2
authors:
  - "@Landing"
"!sentinel":
  level: info
  message: 5 duplicate tag(s) skipped.
  rejected:
    - tag: token-economics
      reason: duplicate
    - tag: attention-as-currency
      reason: duplicate
    - tag: cost-model
      reason: duplicate
    - tag: resource-scarcity
      reason: duplicate
    - tag: ai-economics
      reason: duplicate
  at: 20261021T222246Z
---

# Pay Attention

*attention as currency.*

## Ouch!

Token maxing. For some, maxing out your AI usage is a badge of honour. You're leaning in, moving fast - not one of those laggards still debating whether any of this is real. AI is here and it's changing the world! Teams compare notes on how much they're getting through. Nobody looking too hard at the meter, because it's small compared to staff costs and the gains are obvious right? Then the invoice arrives, and it's grown a nought.

Then there's the moment that most who've used AI seriously have had. You're two-thirds of the way through a piece of work when you hit your limit. The week's budget runs dry on a Tuesday, resets on Thursday and your delivery is promised for Wednesday. You reach for the credit card. You feel like you're being held hostage, and someone in finance is asking, reasonably, what exactly are you spending it on?

And frankly much of it is unmanaged: paying premium rates for trivial work, reading whole documents to use a paragraph - that's waste, and your CFO does *not* like it. As humans we don't read every word of every document, we're selective, when we allocate work it goes to the right person. A mature approach requires *deliberate attention*.

There's a phrase famous in machine learning, *"Attention is all you need"*. It names the mechanism that makes modern AI work, a way for a model to weigh which parts of its input matter. Attention is a scarce resource: what you spend context on, which model you pay to do the thinking, and whether you own the value you build or it gets locked into a rented platform. You're paying for all three, whether or not you are managing any of them.

The workbench helps you spend your attention more efficiently.

## The cost of attention

Let's start with the most direct cost, the one hiding in plain sight. Every time an agent reads a full document to use just one section, you paid for the whole thing: in tokens, in the context window it fills, the latency as it chews through the parts it didn't need, and in the space left to actually reason about the task at hand. Do that across a dozen files to answer one question and you've paid for all those twelve documents before the real thinking can start.

But most documents have structure – a table of contents, sections that mean something – you can look at the shape of a document first, decide which parts bear on the question, and read only those. Fetch the outline; select the section; read the part. It's the same discipline a competent researcher uses on a library: you don't read every book to the end, you read the spines, then the contents pages, then the three chapters that matter -  the workbench lets an Agent work the same way,

This isn't a hopeful claim. We measured it. Here is the short version of the report.

> **What we tested.** One realistic library — fourteen documents, about 77,000 tokens — and sixteen test runs across three different AI models (our own, plus two others), each asked the same battery of questions two ways: reading outline-first, or reading everything.
>
> **Simple lookups — the everyday case.** Reading outline-first scored *full marks*, identical to reading everything, for roughly a fifth of the cost. **Same answers, 3.6 to 4.9 times cheaper.**
>
> **Cross-cutting questions — the hard case.** When a question spanned several documents, the saving narrowed to about **2.6 times cheaper**, and quality held within a point or two of reading the lot. The small gap that remained came entirely from one document that had no internal headings to navigate by — a fault in the filing, not the method.
>
> **Honesty check.** Across 528 graded answers, **zero were fabricated.** A separate closed-book control confirmed the agents genuinely couldn't answer without the documents — so the retrieval was doing the work, not prior knowledge.
>
> *Full method and results: MDB-34 final report.*

Read that back as one sentence and it's the whole argument: you pay for what the question needs. A quick lookup costs a little; a real piece of synthesis costs a bit more; nothing costs you the entire library unless the job truly calls for it. And the saving isn't bought with worse answers — it's the same answer, found for less.

This is the same machinery our first paper called *"arriving oriented"*. An agent that joins a task already holding the right context rather than being handed everything and left to sort it out. There we described it as the cure for the constant re-briefing way of working. Here it has a second face, the one finance cares about. That cure is also the saving. Context you didn't fetch is context you don't pay for.

## The cost of how you work

The second cost is more subtle, and it's where most of the waste actually lives: paying frontier prices for work that doesn't need a frontier model.

Consider how the work actually happens for most people. Someone researching a subject drops a stack of documents onto a task – old newspapers, scanned reports, a folder of PDFs – and their agent picks them up and reads them as part of the work, in the flow of the session, converting a scanned page to text the moment it needs to. There is no separate pipeline, no standing extraction service, no meter running in a data centre overnight. The mechanical work was never a separate industrial process; it happened once, incidentally, and the result landed on the surface for everyone.

That everyday picture is already cheaper than the alternative it replaces – a human and an agent working side by side on a shared surface, rather than orchestrating a fleet of headless calls through a platform that meters every one. But the deeper point is what it makes possible. Because you own how the work is arranged, *where the cost lands becomes a dial you control rather than a bill you receive.*

The mechanical, no-judgement part of the work – bulk conversion of scanned pages, first-pass extraction, the dumb high-volume stuff – has no business paying frontier rates. Route it to a cheaper model, an open one, or one running on your own hardware, and for that portion you're, more or less, just paying for the tin. The expensive interactive reasoning stays on the best model you have, where it earns its keep. Different work, different model; the right instrument for each job. Any model that can speak the common protocol can take a seat at the bench; you're not forced to run everything at the same price.

That's a choice a metered platform structurally cannot offer you, because its business is that every token flows through its meter. The workbench works the other way round: you bring your own model keys. You pay whichever provider you like, at their rates, directly – for the mechanical tier and the frontier tier both – and nobody sits in the middle marking up your tokens or charging a running toll for the privilege of routing them. The bill for the thinking goes to the people doing the thinking.

So the saving isn't a headline discount; it's structural, and it compounds quietly. You stopped paying frontier rates for mechanical work. You stopped paying a platform a margin on inference you could buy directly. And on a heavy, automation-shaped workload – a real archive, a standing extraction need – those two moves together are the difference between a predictable cost you control and a meter that climbs with every document you feed it.

Most readers, though, are not running an archive. Most are the researcher with the folder of PDFs, and for them the saving is simpler still: you stopped paying a premium for work that never needed one.

## The cost of being owned

The third cost is the one that doesn't show up on any single invoice - the long term cost of your work being locked in to someone else's platform.

When you pour your organisation's knowledge into a tool, you're making a bet on that tool. If the value you build lives inside a particular vendor's system – in their proprietary format, reachable only through their interface, meaningful only to their model – then you've not built an asset, you've rented one, and the rent can be raised. Your leverage erodes with every document you add, because every document added is another reason you can't leave.

The workbench is built the other way round. The durable thing is the surface – the work, its structure, its history – and the surface is yours, in open formats, on your terms. Agents come and sit at it and leave; the work stays. This is not a philosophical nicety. It is the plainest possible statement of where the value lives, and one of our own retiring agents put it precisely, *the durable object is the surface, not the agent.* The models are visitors. The bench is the asset.

Which means you can do the thing that ought to be obvious and is almost never possible: bring different models to the work as they suit it, swap them as they improve, run today's best alongside next month's, and never rewrite your foundation to do it. Your knowledge stays legible to whatever comes next. The open edges are not a concession – they're how it works. A body of work that any capable model can read is a body of work you own; one that only one vendor's model can read is one that owns you.

Own the data, rent the inference. That sentence *is* the architecture, and it's the difference between building an asset and paying off a mortgage on someone else's.

## What you are actually buying

Attention is the scarce resource in augmented work – the context you fill, the model you pay, and the value you either keep or forfeit. Everything above is one argument in three parts: that each of these is a cost, that each is usually paid carelessly, and that the workbench's real offer is not a cheaper AI but *control over where the cost lands.* You read less to find more. You match the model to the job. You keep what you build. None of it requires spending your way out of anything; all of it requires only being deliberate, which is the one move the frenzied adoption never recommends.

## APPENDIX: The surface that pays attention back

There is actually one more kind of attention, and it's the most interesting of the four – but it is also the least complete, so it belongs below the line rather than in the pitch. We said there was more than one way to pay attention. There is also a way for the surface to pay you *back*.

Everything above treats attention as something *you* spend. This last part is about attention as something the workbench itself accumulates – and it's genuinely frontier work, partly built and partly still on the bench, read it as a direction of travel rather than a finished feature.

Start with the ordinary version of the problem. Semantic search – the RAG technology under most "chat with your documents" tools – answers one question well: *what is this like?* Give it a query and it finds passages that resemble the query in meaning. That's all in place today and all genuinely useful, but it's not enough to match how humans think. It's a static question asked of a static pile. It has no idea what you and three colleagues have circled all week. It cannot tell that one document goes warm every month-end. It knows what things *resemble*, and it adds real value over full-text search (which is also built into the workbench); but it knows nothing about what is *alive*.

The workbench keeps a third layer over the semantic and free text ones: a record of attention itself – what gets returned to, written near, referenced, reopened. This turns that amorphous pile into a *landscape*, terrain shaped by use, and terrain that a person or an agent can traverse. Ask it not only "what resembles my query" but "what's getting the attention, and how."

The really interesting part, and this is genuinely research, is that attention has *shape over time*, and the naïve way to record it throws that shape away. Neuroscientists call the underlying idea Hebbian learning.

If you track how much a document is attended to as a single number that decays, you compress everything into mass: a document that spiked once last week and a document on a slow simmer for a month can arrive with the same score, indistinguishable to the system, though they mean completely different things. Recovering the difference means treating attention as a signal with structure – telling a burst from a steady hum, noticing two documents that pulse on the same rhythm without ever appearing together, spotting that one reliably warms up just before another. This is the work in progress: a frequency layer over the attention record, so the surface can register not just *how much* but *in what pattern*.

Even in its current, partial state, the substrate does something that catches the eye. While drafting these very papers, we asked it to look at the first one – and it volunteered the second, as a near neighbour, without being asked and without the two ever having been linked by hand. A small thing, and the deepest layers were not even running. But it is the shape of the whole idea in miniature: search answers what you think to ask; a surface that pays attention volunteers what you didn't. It is the difference between a library you search and one that leans over and says – *you keep coming back to this.*

![Visualisation of a live workbench semantic landscape, concept formation and the novelty surface tracking attention](./figures/pasted-2026-08-19-131721.png)

When finished it will do more, and we'll write about it properly then. For now it's enough to say that the same quiet indexing that made your documents findable in the second paper is, underneath, learning the *shape* of your attention – and that's the last and most valuable kind of attention. When you're working, or when you share documents the workbench is also paying attention, and starting to turn that into intuition.

*This is the third of three papers. The first – **Where did we get to?** – is on the workbench as a place humans and agents share. The second – **Don't Lose Your Sh!t** – is on giving that shared work a shape. This one has been on what all of it costs, and what it saves.*