JA Westenberg on agency: direction is the new bottleneck

Key ideas

• An agent acts on someone else's behalf. The principal is the person who decides what the agent should do and answers for the result.
• AI agents make doing things cheap. They don't make it any cheaper to know what you want.
• A clear instruction multiplies your output. A vague one multiplies the drafts you have to read and throw away.
• Four jobs stay with the person: wanting a specific outcome, choosing what to give up, judging the result, and owning it.
• Taste comes from doing the work by hand, and it fades if you stop.
• Measure a week by which outcomes moved, not by how many tasks the agents closed.

JA Westenberg’s argument is that when AI agents make doing things nearly free, the scarce skill is deciding what you want. He borrows two words from the old meaning of agent: the agent acts, and the principal is the person the agent acts for. He says the principal has four jobs left (wanting, choosing, judging and owning the result) and gives six habits for practicing them. These are my notes on his 8-minute video.

It’s an opinion piece. He cites no studies, so the numbers below (ten times the output, twenty years of habit) are his figures of speech, not measurements.

Agent and principal

He starts with the plain meaning of the word. A lawyer, a real estate broker and a proxy are all agents: each one does something for another person. That other person is the principal.

His claim is that software agents are about to make everyone a principal. They’ll book the flights, write the code and clear the inbox. The cost of doing those things falls. The cost of being the principal doesn’t, because the principal still has to decide what needs doing, look at the result without fooling themselves, and take the blame when it goes wrong. He thinks that gets harder as the tools get better: when almost anything is one instruction away, you hit the question of what you’re trying to do sooner, and more rides on the answer.

Why the habit is weak

Most people, he says, have spent their working lives as someone else’s agent. A boss wanted a spreadsheet, a ticket was assigned, a calendar said where to be. School trained it first, with a syllabus handed over at the start. Then a feed picked the next video before you’d noticed you were bored.

Each of those took a small decision away, and each time it felt like relief, because choosing is hard and being told is easy. After two decades of it, in his telling, people are very good at reacting and out of practice at starting. Give a person like that ten capable agents and not much happens. They have nothing to point them at.

What a vague instruction costs

You can ask the machine what to work on, what to write or who to email, and it’ll answer every time. He says the answer will be fine, and that’s the problem: it’s roughly the average of what everyone in your position has done, so following it makes you average faster.

He also argues that standing still isn’t an option. A clear instruction gets you far more output for the same effort. A vague one gets you the same multiple of drafts, options and plans, and you have to read and sort them all. The gap then compounds. Someone who knows what they want gets a result, learns from it and wants something sharper next week. Someone who waits gets a pile of output, feels busy, and learns nothing about what they want.

The four jobs of a principal

The talk describes the fourth job without giving it a one-word name, so “owning” is my label.

On judging he makes two points that go together. Taste only comes from having done the work yourself, slowly, until you’re good at it. And it expires: he says that after about two years away from the hands-on work, you stop being able to see what good looks like now.

Six habits

1. Write the want before the prompt. One sentence, by hand, before opening the agent: by the end of today, I want this. If you can’t write it, he says the problem isn’t the tool.
2. Say what done looks like. An agent will keep running as long as you let it, so write the finish line first. His examples are three profiles with a one-line reason each matches, or a draft under 800 words. If you can’t describe the finish, you’ve decided to be busy, not what you want.
3. Keep a wants list, not a to-do list. Agents can produce to-do lists all day. The wants list is short, covers this quarter, and is in your own words. Read it each morning, and drop any task that doesn’t trace back to a line on it.
4. Take a position. A model hedges, in his account, because it learned from a world that disagrees with itself. The principal has to decide and commit. If you’re wrong you find out quickly, which he counts as a benefit.
5. Do the work you’d never delegate. Everyone has one task they’d still do with no tools: the sales call, the first draft, the pricing decision, the hard conversation. He says that’s where your agency is still in use, so protect it.
6. Review the outcome, not the activity. At the end of the week, look at the wants list and ask what moved and in which direction. Ignore how many agents ran or how many tasks closed.

Where to start

The cheapest test of the whole idea is habit one. Before the next agent session, write the sentence: by the end of today, I want this. Then write the finish line under it. If both come easily, the rest of his list is bookkeeping. If they don’t, that tells you where the work is.

His closing point is that this is good news for people without a technical background. In the old picture the limit was capacity: hours, hires and skills. He argues that capacity is now close to free, so the limit is direction, and deciding what you want is something everyone could do before years of being told what to do trained it out of them.

Key terms

Agent

Someone, or now some software, that does things on another person's behalf. The word comes from the Latin agere, to do or to act.

Principal

The person an agent acts for. The principal picks the goal, judges the work and takes the blame.

Personal agency

His name for the habit of deciding to do something and starting, without waiting for permission.

Want

A specific outcome you can state in plain words, with a date. He contrasts it with a wish such as growing the business.

Wants list

A short list of the outcomes you're after this quarter, in your own words. Every task you hand an agent should trace back to a line on it.

Finish line

A description of what done looks like, written before the task is assigned.

Quotes from the talk