Skills vs. AgentsThe Breakdown

Everyone in this room is renting the same intelligence. Same weights. Same context window. Same API, the same week, at the same price. The spread in what people get out of it is enormous — and it is not talent, and it is not access.

Same model, different results

Some people have an agent that drafts their week while they sleep. Most people have a very expensive autocomplete that they argue with.

The gap is that almost nobody can say precisely what the parts of an agent are, which parts they are renting, and which parts are theirs to build.

So let us define the thing carefully, part by part. Then a story about an airplane — and by then you will already know why.

What an agent is

An agent is an LLM running in a loop, calling tools until it completes a task and returns a result.

Take it apart word by word.

In a loop. This is the whole thing. A chatbot answers once. An agent keeps going until it is done.

Calling tools. It can touch things. Files, a terminal, a browser, your calendar.

Until it completes a task. It decides when it is finished. Nobody is stepping it forward.

Four boxes, and one of them is a for-loop. We have been calling this magic for two years. It is a while-loop with a language model in it.

The three properties

An agent, as distinct from a model you chat with, has three properties.

1. It can perceive its environment. It can read the file, list the directory, load the page, check the calendar. It is not working from your description of the world; it is working from the world.
2. It can make autonomous decisions. Given a goal, it chooses the next step. You did not specify the step. That is the part that feels uncomfortable, and it is also the entire value.
3. It can act with little or no human intervention. It writes the file, sends the request, runs the script. Perception without action is a chatbot with a browser. Action is what makes it an agent.

Notice that none of the three is “be smart.” Smart is table stakes, and it is getting cheaper every quarter.

The six things an agent needs

Properties describe what an agent is. This is what it needs in order to be useful on real work rather than in a demo.

1. An environment where it can take actions. A world with a surface it can touch. No tools, no agency.
2. A way to manage growing context over long runs. Every step adds to the transcript. Real work overruns any window. Something has to decide what stays loaded and what gets dropped.
3. A connection to your data. Otherwise it knows what everyone else knows, which is worth exactly what everyone else’s answers are worth.
4. The ability to parallelize. Some work fans out. One agent reading forty documents in sequence is a bottleneck you invented.
5. A human in the loop. Not for everything. For the irreversible things: the send, the delete, the deploy, the payment.
6. A way to improve over time. Here is the trap. The agent does not improve. It wakes up every morning with total amnesia and no memory that you exist. The model improves when a lab ships, which is neither yours nor on your schedule.

Hold on to number six. Everything else in this document is an answer to it.

The three parts

An agent is not a list of three ingredients sitting side by side. It is a nesting, and the nesting is the point.

Read it from the outside in.

The environment is the world the agent runs in. Your files, a terminal, a browser, whatever data you have connected to it. The agent does not sit outside this looking in through a window. It runs inside it, the way a program runs inside an operating system. That is what makes it an agent rather than a chat box: it is already in the room with your work.

The agent is the whole running thing. It is not a component you can point at. It is the name for all of this, in motion.

Inside the agent sits the harness. This is the piece people skip, and it is the piece doing the actual orchestration. It holds two things and coordinates between them, turn after turn, until the job is done.

The model is the reasoning, and the context is everything the model can see when it reasons. The harness decides, on every single turn, what goes in front of the model and what the model is allowed to do about it. Neither the model nor the context does anything on its own. The harness is what makes them into a loop.

A harness orchestrating a model over a context, running inside an environment. That is an agent.

One thing to notice now, because it changes what the rest of this document is about: skills are not a fourth box. They live inside Context. A skill file is a kind of context, and we will get to exactly which kind shortly.

Take the three in turn, because which of them is yours turns out to be the only question that matters.

Two of those rows are going to zero. One of them compounds. Almost every argument you will hear about AI is about the first two.

The harness

The harness is the program that runs the loop. Claude Code, Codex, Cursor, any agent CLI. It is the thing you actually launch.

The clearest way to define it: the model thinks, and the harness does. The model produces a decision about what should happen next. It cannot open a file, wait three seconds, retry a failed call, or stop and ask you. The harness does all of that, then hands the result back to the model and asks again.

Go back to the six things an agent needs. The harness is what supplies five of them.

That table is the argument of this whole document in one image.

The harness orchestrates. It decides which of your files are open on the desk at this moment, which tools are within reach, when to stop and ask permission, and when the job is done. It is genuinely sophisticated engineering, and you get all of it for free. Every harness on the market is converging on the same feature set, and they are racing each other to zero.

Which means the harness is not where your advantage lives. Everyone has one, they all do roughly the same job, and the one you pick today is not the one you will be running in eighteen months. You can swap it out on a Tuesday and lose nothing at all, provided your knowledge was never welded into it.

The model

The model is the reasoning. It is the thing that reads a messy situation and works out what to do about it.

What it genuinely is: the best general reader and writer ever built. Give it a vague request and it will infer what you meant. Give it a page of prose and it will extract the structure. Give it tone, taste, judgment, ambiguity, and it performs at a level that would have been science fiction five years ago.

What it is not: yours. It is rented, it is a commodity, and it is getting cheaper by the quarter. Everyone reading this has access to the same one.

And it has one hard limit that shapes everything else. The model knows what everyone knows. It has read the public internet. It has not read your inbox, sat in your meetings, or watched you handle the client who goes quiet when he is unhappy. On the day you need it most, it knows nothing about your situation.

That is not a flaw to be fixed by a better model. A better model reads faster and reasons harder over whatever you give it. Which brings us to the part that is actually yours.

The context

Context is everything the model can see when it makes a decision.

This is the part almost everyone treats as an afterthought, and it is the only part of the system that is genuinely yours.

Start with the constraint on the human side. You hold about seven things in your head at once. Seven plus or minus two, the most cited number in cognitive psychology. It is why phone numbers were seven digits, and why you forget the eighth thing on the grocery list. Every institution humanity has built is a workaround for that number: every checklist, every org chart, every filing cabinet, every stand-up meeting.

An agent holds roughly a thousand pages at once. Every email you sent last year, printed and stacked on the desk, open simultaneously, searchable in seconds.

That sounds like abundance. Run the number the other way and it is not. A thousand pages is enormous next to seven and it is nothing next to a life. Your working life is not a stack of paper; it is a library. Every decision and the reason behind it. Every conversation with everyone you know.

So the question that decides whether your agent is a genius or a goldfish is not how big is the window. It is:

Who decides which pages are open?

And to answer that you have to notice that “context” is not one thing either.

Three kinds of context

Psychologists split human long-term memory into kinds, and the split turns out to be exactly the one we need. Machines have the same three, and almost everyone building with agents is only feeding two of them.

Episodic context: what happened

The record of events, in order, with time attached.

Your sent mail. Meeting transcripts. The thread where the scope changed. Last week’s agent run and what it did. Episodic context is a history, and its defining feature is that it is about particular moments: this person, on this date, said this thing.

It is the easiest kind to accumulate, because you generate it whether you want to or not. You are sitting on ten years of it in an inbox right now, unindexed, doing nothing. That is a moat lying on the ground.

It is also the kind that goes stale fastest and the kind most likely to mislead. A decision from March that was reversed in June is still sitting there, perfectly preserved, ready to be retrieved with total confidence.

Semantic context: what is true

Facts, stripped of when you learned them.

Your pricing. The architecture. The eligibility rules. Who owns what. Semantic context is a reference library, and it is what almost everyone means when they say “we’re doing RAG”: chunk the documents, embed them, retrieve the relevant ones, put them in the prompt.

This works, and it is worth doing. It is also where most teams stop, and it is why so many internal AI projects feel like a slightly better search box. Because retrieval answers what is true, and almost no real work is blocked on that.

Ask yourself honestly how often you are stuck because you could not find a fact, versus stuck because the thing is just fiddly and has eleven steps and two exceptions and only you know the order.

Procedural context: how it is done

Not what happened. Not what is true. How the thing gets done.

The order of the steps. What to check before you start. The exception that only comes up twice a year. The thing you do differently for that one client. The point where you stop and ask a human instead of proceeding.

This is procedural memory, and in humans it is the strangest kind, because it is the kind you cannot easily say out loud. You can recite facts. You can recount events. But ask an expert how they do the thing they are best at and you get “well, it depends,” followed by twenty minutes of watching them do it while they narrate badly.

That is the crux. Your episodic context is captured whether you like it or not. Your semantic context is mostly written down somewhere. Your procedural context is not written down anywhere, because the only place it has ever needed to exist is inside your head.

And a procedure inside a head has three properties that should worry you. It cannot be delegated except by standing over someone’s shoulder for a month. It cannot be executed while you sleep, because it needs you. And it cannot be reached in an emergency by anyone but you.

A skill file is procedural context, written down and made executable. That is the whole idea. Before defining one properly — what it costs to leave procedural memory where it usually lives.

The Gust Lock

Wright Field, Ohio. October 30th, 1935.

A man named Ployer Peter Hill climbs into an airplane. He had turned forty-one six days earlier.

He is the Chief of the Flying Branch at Wright Field, which means that on this morning, in this country, there is no better pilot. If you wanted to know whether an aircraft was safe, you gave it to Pete Hill. That was the job. That was the whole job.

The airplane is a Boeing Model 299. Four engines, fifteen tons, machine guns bristling out of every surface. A reporter at the Seattle Times had taken one look at it and called it a flying fortress, and the name stuck so hard that Boeing trademarked it before it had ever fired a shot in anger.

Boeing had built it with their own money. No contract. They bet the company on it. And it was winning: at the fly-off it embarrassed everything Douglas and Martin brought, and Army procurement officers were already talking about buying sixty-five of them before the competition had even finished.

This morning is a routine climb test. Find out how high it goes and how fast it gets there. Five men board. Hill in the left seat. Lieutenant Donald Putt beside him. Behind them, standing in the aisle, Leslie Tower, Boeing’s own chief test pilot, the man who flew this airplane on its very first flight, riding along today as an advisor. Two more in back: a test observer and a mechanic.

It rolls. It lifts. It climbs to about two hundred feet.

And then it stalls, rolls over, and goes into the ground. It burns.

Three men get out. Pete Hill dies a few hours later. Leslie Tower is pulled from the fire with burns across his face, his right arm and both legs. He holds on for twenty days. He dies on the nineteenth of November.

There was nothing wrong with that airplane

The engines were fine. The airframe was fine. The design was so fine that twelve thousand seven hundred of them get built, and they go on to drop more bombs than any aircraft in the Second World War.

What killed Pete Hill was a gust lock.

It is a pin. When an airplane is parked, the wind pushes on the elevator and the rudder and slams them around, so you lock them still on the ground. Before you fly, you unlock them.

Nobody unlocked them.

The most experienced test pilot in the United States Army took off with his controls pinned, and there was nothing he could do, because the thing he needed to do he needed to have done four minutes earlier, on the ground.

It was his first flight in the type.

The detail that matters here

Leslie Tower, standing behind the pilots, the man who had built this airplane’s flight envelope with his own hands, saw it.

He saw the locks were still engaged. He knew exactly what was about to happen and exactly which lever fixed it. He moved for it. He could not reach it in time.

Read that against the last section. Tower had the procedural context. He was probably the only man alive who had it completely. And it was stored in the one place procedural memory is always stored: inside a human head, from which it could not be transferred to where it was needed inside of four seconds.

That is not a story about a careless crew. It is a story about what happens when the procedure lives in a person instead of in the aircraft.

What they did next

The crash disqualified the Model 299. The Army cancelled the order and bought a hundred and thirty-three twin-engine Douglas bombers instead. Boeing nearly went under.

The verdict, as a newspaper put it at the time, was that the Model 299 was “too much airplane for one man to fly.”

Sit with that. Not a bad airplane. Not a bad pilot. A gap between how much machine there was and how much a single human being can hold in their head at one time. Seven plus or minus two, in a cockpit with four engines.

So here is what the Army did.

They did not go looking for a better pilot. Think about how important that is. The obvious response to “our best man crashed it” is “get a better man,” and there was no better man. Pete Hill was the better man. That road was closed. They did not simplify the airplane either, because they needed the airplane.

What they did instead was sit down a group of pilots, and those pilots wrote out, on a card, the things you do before you start the engines. Before you take off. In the air. Before you land. After you land.

A list. In English. Of the obvious things.

Every pilot already knew every item on that card. That was never the problem. The problem was that knowing it in your head, on a busy morning, at forty-one years old, six days after your birthday, with four engines running and a schedule to keep, is not the same as having it written down in front of you.

They put the card in the airplane. Every commercial flight you have ever taken began with two people reading a list out loud to each other. Every single one. That is Pete Hill’s card. It is still in the cockpit. It never left.

The answer to a more capable machine was never a more capable operator. It was procedural context, extracted from a head and put where the work happens.

That was settled in 1935. We are rediscovering it now, at enormous expense, with language models.

Why skills are the part that matters

Come back to the six things an agent needs, and to number six, the one the harness could not supply: a way to improve over time.

Now we can answer it.

The agent does not improve. The model improves on a schedule that is not yours. The harness improves and hands you the improvement for free. The only thing that improves on your schedule is your context, and the only part of your context that encodes judgment is the procedural part.

That is what a skill file is. A page of English describing one job, clearly enough that a competent stranger could execute it, stored as a file, loaded on demand, reused forever.

A skill file is a checklist that can fly the plane.

That is the only difference between Pete Hill’s card and yours. His needed a pilot. Yours does not.

A runtime and a procedure

An agent is capability. A skill is judgment. Capability is rented and getting cheaper. Judgment is yours and getting rarer.

The architectural version of the same claim: thin cockpit, fat checklist. Put almost nothing in the harness and almost everything in the skills. You will switch tools twice next year, everybody will. If your knowledge lives in the harness, every switch is a lobotomy. If it lives in flat files in a folder you own, a switch is an afternoon.

Whose repo is it in?

One more thing before the how-to, because a skill file is not a document. It is a piece of your cognition: how you do the thing, taken out of your head, written down, and made executable.

There is a proposal manager at a mid-size contractor a few miles from here. Call her Renee. Over three years she teaches her agent forty skills. How to read an RFP and pull out what they are actually scoring. How to spot a past-performance gap early enough to fix it. How to write a pink team draft that survives the red team. How to tell, by Tuesday, whether a bid is worth the bid-and-proposal money.

Forty files. Three years of judgment nobody taught her, sitting on a disk.

Version one. Those files are in Renee’s repo. She moves to another contractor and they go with her. Day one, she is operating with three years of compounded judgment on tap. That is ownership.

Version two. Those files are on the company share drive, under the company’s IT policy. She leaves with nothing. The company keeps running her judgment without her, on every bid, forever, and her name is nowhere in it.

Same forty files. Same Renee. One variable.

She did not have a career. She had an extraction.

An enormous number of us carry institutional knowledge that lives in somebody else’s system.

How to write a skill

Here is a real one. Not an excerpt, the whole file.

skills/bootcamp-applicant-triage.md

When a bootcamp application comes in:
1. Check the applicant is 8 to 19 on the start date. Flag anything outside.
2. Route to the Nigeria track or the USA track. Start dates differ.
3. If guardian consent is missing, draft the note to the parent. Do not send it.
4. Add them to the cohort sheet: math, coding or design.

If an application is incomplete, hold it in Pending. Never reject anyone on my behalf.

That is it. A page of English. No code, no framework, no schema.

It is one job, written down. A volunteer starting Monday could follow it without asking a single question.

And that is the test. The Monday test: if someone starting Monday could run it without calling you, an agent can run it.

What every good skill file has

• One job. A skill that does four jobs is four skills. The single most common mistake is writing a file that describes a whole role.
• A trigger. Say when to use it, in its own opening line. That is how the harness finds it among fifty others.
• The steps, in order. Numbered. Boring. The order is information.
• The exceptions. This is the part that separates a real skill file from a memo, and it is where most of the value lives. Guardian consent missing? Draft it, do not send it. Every one of those is a gust lock: a step completely obvious to you, completely invisible to your agent, and fatal when skipped.
• A stopping rule. What it must never do without you. Never reject anyone on my behalf. Irreversible actions get a human.
• Where things live. Which folder, which sheet, what the output looks like.

Write it by getting it wrong

You do not author a skill file. You correct one into existence.

Explain the job to your agent in plain English, the way you would explain it to a smart friend on their first day. Then run it and let it get it wrong. It will get it wrong, and that is not failure, that is the process: it is discovering the gust locks you did not know you were carrying.

Every correction, every “oh, and also,” goes into the file. After four or five passes the file stops being wrong. It is now an employee.

Everyone at Wright Field in 1935 knew about gust locks too. Knowing it was never the deliverable. Writing it down is the deliverable.

Know which half is which

The most useful practical idea here, and the one that prevents most agent failures. There are exactly two places computation can happen, and confusing them causes nearly every problem.

Go back to the application file. Write the follow-up to a parent whose consent form is missing, warmly, without making them feel accused is latent space. That is taste, and the model is genuinely good at it.

Tell me which of three hundred applicants are eligible on a start date under two different track rules is arithmetic across a table. Ask a model to do that in its head and it will hand you a confident, wrong list. It has to write code and keep the answer somewhere countable.

The model fails exactly where we fail. Your skill file carries the judgment; the code it calls does the arithmetic. Do not make the model do arithmetic and do not make the database do judgment.

The “it is just a memo” objection

Somebody always says it. In 1953 Grace Hopper proposed that you should be able to tell a computer what to do in English. Her management ruled it unfeasible. In her own words:

“Nobody believed that. I had a running compiler and nobody would touch it. They told me computers could only do arithmetic; they could not do programs.”

She built it anyway. It took two years to convince her own company. It shipped as FLOW-MATIC, the first language to use English-like statements instead of symbols, and it became the direct ancestor of COBOL, which ran the world’s payroll for fifty years. She was right, and so completely right that we forgot it was ever controversial.

Markdown is code. If you can write clear instructions in English, you are a programmer. The compiler is a language model.

One honest caveat

A library nobody weeds is not a brain. It is a filing cabinet nobody has opened in two years, except now it answers instantly, with total confidence, and it is out of date.

A bad skill file encodes a bad process forever, and that is the part people miss. You are not writing a note. You are writing something that will execute a thousand times without ever asking whether it still makes sense.

So: where did this come from, what happens when new information contradicts old, and who prunes. Treat it like production infrastructure and it compounds. Treat it like a dumping ground and you get a very confident agent that is wrong in ways nobody can trace.

What to do right now

Five things. Do them and in ninety days you will be operating in a way that looks, from outside, like cheating.

1. Pick a harness — tonight

Not a decision, a coin flip. Claude Code, Codex, Cursor, any agent CLI. Any of them does 99% of this.

Do not spend three weeks choosing. Choosing is procrastination that feels like research. Pick the one your friend uses so you have someone to ask.

2. Start your library — this weekend

Not a grand archive. One folder of markdown files.

Write one page about each project you are working on and each person you work with. On those pages put what you actually know: what you are building, what they care about, what you owe them, what they said last time.

That is episodic and semantic context, and no model on Earth has it, because it exists in exactly one place and that place only holds seven things.

The first time an agent answers using your details instead of the internet’s, something clicks and you do not go back. You do not start with the archive. You start with one page.

3. Write your first skill file

Now the procedural half. Picking it is easy: what is the task you do every week that you hate the most? Expense reports. Meeting notes. The Friday status update. Competitor research.

Explain it, run it, let it get it wrong, fold every correction back into the file. Four passes and it stops being wrong. When you are done, that page is not a document. It is an employee.

4. Wire one recurring job

Take the skill you just wrote and put it on a schedule. Every morning at seven. Every Friday at four.

The first time you wake up to work that finished while you were asleep, something changes permanently. That is the morning you stop being the bottleneck.

5. Never do one-off work

This is the one that separates the people who compound from the people who dabble.

Most people run one task, get a good result, close the window, and throw the context away. Then next month they do the identical thing from scratch.

Do not. At the end of every task, ask the agent to skillify it: turn what it just did into a markdown file you can reuse forever.

If you have to ask for something twice, you failed.

The agent does not improve. The library does.

What ninety days looks like

Week 1. It is a toy and you will feel silly. The library is thin, the first skill is clumsy, you spend longer correcting it than doing it yourself. This is where almost everyone stops.

Week 4. It starts answering with your details instead of the internet’s. Something lands in your inbox before you are awake and you actually read it. You write the third and fourth skill, because the first two held.

Week 12. You stop explaining yourself. The work you dreaded on Sunday night is already done by Monday, and somebody at your job asks if they can borrow the thing you built for yourself.

Flat, flat, flat, then not. Almost everyone stops in week two, which is exactly why the ones who do not feel like they are cheating by week twelve.

And you do not have to be an engineer

The skill file above is from a foundation that teaches free coding, design and maths bootcamps to children aged 8 to 19, staffed entirely by volunteers: a treasurer, a secretary, a VP of skill training.

Not one of them is a programmer. They do not need to be. They need to be able to describe their own job clearly, in English, once.

That is the entire barrier. If you can train a new hire, you can do this.

The card outlived both of them

In 1939, four years after the crash, the War Department needed a name for a new air depot in Utah. They called it Hill Field. Today it is Hill Air Force Base. Twenty thousand people work there, and most have no idea they go to work inside a memorial to a man who died because of a pin somebody forgot to pull.

Pete Hill got a base named after him.

Leslie Tower, who saw it, and reached for it, and could not get there, got nothing. There is no Tower Air Force Base. He was a company test pilot and he died of his burns in a hospital in Dayton and that was that.

But the card outlived both of them. The card is in every cockpit on Earth tonight.

Right now, today, you are carrying around a Leslie Tower. You know how to do something well. It is real expertise, it is genuinely yours, and it lives in exactly one place, which is the inside of your head, where it cannot be reached. Not by your team, not by your agent, not by you on a bad morning.

Every one of us has a card we have not written yet.

You are one page of English away from it being worth something after you leave the room.

Strip away the loop and the model still does six things on its own — for free, in any chat window.

There is a seventh item on every list like this, usually labelled “agents.” It is not a seventh capability. It is what happens when you take the first six and stop asking a human to carry the input and the output by hand.

This is the part you act on.

The Clinic

Case 1 · Organization · Healthcare

Riverside Family Medicine — three physicians, two nurse practitioners, a front desk of four. Small enough that everyone knows everyone, busy enough that nobody has a spare hour.

Ninety minutes a day, back. Not because the clinic bought a better model — because somebody wrote down what Whitney already knew.

The Retailer

Case 2 · Organization · E-commerce

Kettlebell & Co. — a direct-to-consumer fitness brand, one warehouse, six people, most of them doing three jobs each. Here the volume is the problem, not the sensitivity.

“Changed my mind” and “arrived broken” are the same three words of typing and completely different business decisions. That distinction lived in someone’s head. Now it is a file.

The Analyst

Case 3 · Individual · Inside a company

Priya does competitive intelligence for a mid-size B2B software company. The Monday digest takes most of her Sunday night, which she has never told her manager.

She did not file a ticket with IT. She did not wait for a company AI initiative. One page of English, her own laptop, a Tuesday night.

The Operator

Case 4 · Individual · Working for yourself

Marcus runs a solo consulting practice — strategy work for small manufacturers. One-person shop, no ops team, because he is the ops team.

The agent does not improve. The library does.

Two tracks out of this room

Company or individual, three people or three thousand, the pattern repeats. Find your column.

Right now you are carrying around something you know how to do well, that lives in exactly one place, where it cannot be reached — not by your team, not by your agent, not by you on a bad morning.

You are one page of English away from it being worth something after you leave the room.

Go write the card.