01 · What is agency?

What is agency?

Not motivation. Not autonomy alone. Not efficiency.

Agency is how much of the future you can actually move from where you stand.

It lives in a loop: what you want, what you do, what actually happened, what you change next. The tighter you close it, the more you can cause.

The strict version: How effectively an actor closes the loop from intent → intervention → evidence → adaptation. Not the same as power (what consequential transformations are available), capability (can it do X), authority (is it permitted), influence (how much it changes others) or control (how reliably it steers). You can have huge power and poor agency, or high agency in a tiny domain.

02 · Three futures

Here's you. Now.

Physicists draw a moment like this: a point, the past trailing below it, and above it a cone of every future that can still be reached from here.

Every point of light is a future. A different job, a finished book, a company, a Tuesday that goes differently.

Most people think there's one cone. There are three.

Inside the big one: the futures you believe are reachable. Your map. You can't aim at a future you can't picture.

And inside that, the small bright one: what you can actually cause, with the skills, permissions, money, people and hours you have right now.

That's your agency. Not what you know. Not what you could imagine. What you can move.

03 · The future you can't see

Somewhere out there is Z.

Say reality has a path from here to Z. Somebody knows a trick you don't. Z is real. It just isn't on your map.

So “that's impossible” is almost never true. The honest sentence is “I don't know a path to that yet.”

The first ends the conversation. The second is a to-do: search, ask someone, find out, run a test.

04 · What AI actually did

It blew your map wide open.

Ask a model anything and you can suddenly see routes you never knew existed. Z is on the map now.

Your bright cone grew too, but less, and only where you acted on what you saw.

The difference is the gap.

It's why you can feel more capable than ever and still be stuck on the same things as last year.

05 · Five ways to widen it

“I can do more now” hides five different things.

A route.

Someone shows you a path you couldn't see.

A capability.

You can do the thing, or you have a tool that does it.

A permission.

You're allowed to act: the role, the sign-off, the license.

Resources.

The money, time or materials the path costs.

Coordination.

The other people moving with you.

The first widens what you can see. The other four widen what you can cause. They are not the same fix.

06 · Ten things at once

Agency isn't one number.

It depends on ten things at once. Each petal is one of them.

And AI affects each of them differently.

• expands grew, a lot
• mixed grew in one way, shrank in another
• flat didn't move
• erodes got worse by default

1 of 10

Intent erodes

Do you know what you want, and does it hold still long enough to act on?

A model can lay out a dozen futures in a minute. It can't want one for you, and the flood of new options makes the target move more, not less.

The move: Pick the target yourself and write it down. Then let the machines work toward it.

2 of 10

Epistemic expands

Can you see what's actually happening, and how wrong is your picture of it?

Expands the most. You can see more routes, more options and more of how things work than anyone could a few years ago. It also hands you confident pictures that are wrong.

The move: Check the picture against the real source before you act on it.

3 of 10

Intervention expands

What can you actually do to reality, how precisely, and how badly can it go?

What you can actually do exploded: write, code, research, design, analyze, operate. Precision didn't come with it, and neither did knowing which action moves the result.

The move: Find the one action that changes the outcome, and spend the new reach there.

4 of 10

Resources mixed

What can you spend, and which of it is scarcest?

Labor and compute got cheap: work that needed a hire now costs a few dollars. Money, relationships and trust didn't move. Your attention became the scarcest thing you have, and AI spends it by default.

The move: Measure what one minute of your attention causes, and guard it.

5 of 10

Coordination mixed

How many things can move at once without routing everything through you?

You can run ten things at once now. By default all ten route through you, and you become the bottleneck.

The move: Give each agent the intent, the context and the limits up front, and have it bring you only the exceptions.

6 of 10

Institutional flat

What are you allowed to do, and who recognizes it?

Unchanged. No model can give you permission, standing, a contract or a license. Agents raise the stakes: they will act on permission you never gave.

The move: Grant authority on purpose, in writing, and keep what you want separate from what you've allowed.

7 of 10

Temporal mixed

How fast can you act, and how long does what you do last?

Speed went up by orders of magnitude. Staying power didn't: the model forgets when the conversation ends, and most of what it produces evaporates.

The move: Store what worked somewhere that lasts: a method, a file, a checklist, a receipt.

8 of 10

Feedback erodes

When you act, how quickly do you find out what it did, and do you learn from it?

More output, less finding out what it did. A draft isn't an outcome, and a finished task isn't proof that anything changed.

The move: Build agents that read the real result back, and look before you call it done.

9 of 10

Reliability erodes

When you try, how often does it actually happen the way you meant?

Long chains break more often than they look like they will. Ten steps that each work 95% of the time fail four times in ten.

The move: Shorten the chain, and check the work between the steps.

10 of 10

Amplification expands

What does it cost you to do it again, a thousand times?

Doing it again got almost free: copy it, send it, run it a thousand times. Mistakes got the same discount. A bad rule runs a thousand times as easily as a good one.

The move: Make it work once, verified, before you let it run a thousand times.

07 · The weakest link

Your weakest link is your ceiling.

Agency is only as strong as the weakest link in the loop. Cheap cognition strengthened some links enormously: what you can see, what you can do, how fast, how often. It left the rest where they were, or wore them down: what you want, what you're allowed to do, whether you find out what actually happened.

Thinking got almost free. Agency didn't follow automatically.

08 · The gates

Why that book didn't change your life.

Here's everything a great book could do for you, pouring down from the top. Between the page and a changed life there are 8 gates. The light has to get through all of them.

Hand it to an alien.

Give a visitor from another planet every book humans ever wrote. They can't read a word. One gate is shut, so the agency they received is about zero.

Now give them a translator. Nothing in the books changed. The coupling did. That's what AI models are for most people: the cheapest translator in history.

Now you.

You read the book. You understood it. You even agreed with it. Then: could you actually do it? Were you allowed? Did you?

A strong page times a shut door is a shut door.

09 · Long chains break

95% ten times is 60%.

Ten steps that each work 95% of the time give you a process that works 60% of the time. Chains of agents multiply their failure rates, not their intelligence.

The move: shorten the chain, and check the work between the steps.

10 · Where agency is stored

A book is frozen agency.

Somebody spends twenty years figuring something out, then writes it down. The book holds the potential to change what someone can do. It can sit on a shelf for ten years and do nothing.

Agency is stored everywhere effort outlasts the moment: attention, writing, money, skills, trust, roles, teams.

That's the whole trick of civilization: somebody pays for the discovery once, and everyone after gets it cheap.

11 · The last row

Every era cut one cost.

Books stored it. Search found it. Phones put it in your pocket. Language models cut the cost of decoding it: the right idea, in your words, at your level. Agents are starting to cut the cost of acting on it.

One row is still empty.

Nobody has cut the cost of keeping it together: holding the goal, remembering what worked, keeping five tools and three people and a month of effort pointed the same way.

That row is still yours. It's where most agency leaks, and where the biggest edge is.

12 · The loop

Ten moves. The last one lights the rocket.

Books, models, mentors: all carriers. What turns them into a changed life is a loop you run:

1. Orient. Form a usable model of "where am I, and what's happening?"
2. Value. Distinguish what I care about and what I'd prefer.
3. Imagine. Represent states that don't exist yet.
4. Causalize. Ask what would have to become true for that.
5. Intervene. Turn one part of the causal frontier into an actual attempt.
6. Observe. Distinguish what happened from what I hoped would happen.
7. Update. Change the model or the intervention rather than defend the story.
8. Acquire decoders. When blocked by ignorance: learn, ask, research, recruit, tool up.
9. Persist. Keep enough state that an interruption doesn't reset the whole process.
10. Recurse. Apply the same process to the limitations of the process itself.

Then you hit a wall.

You don't know how. The paper is in Japanese. The tool has no manual. Most people stop here. So do a lot of AI agents.

The last moves are how you don't: acquire what you're missing, keep your state so an interruption doesn't reset you, and recurse: run the loop on the loop's own limits.

Above that threshold, almost every block turns into another problem you can solve.

13 · Compound

Close the loop. Then close it again.

Every lap leaves something behind: a method, a tool, a receipt, a relationship. The next lap starts further out.

See more.Choose better.Cause more.Compound everything.

14 · Your move

Don't leave with a better map. Leave with a move.

Pick one thing you want that hasn't happened yet. Ask one question:

Can I not see a path, or can I not take it?