# Prompts, Skills, and MCP Will All Disappear. People Only Pay for Outcomes - Author: Rory Cai (https://coiggahou2002.github.io/) - Published: 2026-09-19 (Asia/Shanghai; 2026-09-18T16:34:00.000Z) - Language: en - Canonical: https://coiggahou2002.github.io/blogs/pay-for-outcomes/ - Chinese version: https://coiggahou2002.github.io/zh/blogs/pay-for-outcomes/ Some thinking that came out of recent debates on our team. All the prompts, skills, hooks, and plugins everyone's busy studying today\ are just a transitional phase of AI applications. They won't stick around. What sticks around will be a market:\ a market of agents that deliver nothing but outcomes. Why? People have actually argued about this many times before, in different forms. Let's walk through it. ## 1. If you can outsource anything, why do companies exist? Start with a question: why do companies exist at all? In 1937, a 26-year-old Brit wrote a paper\ answering exactly that.\ His name was Ronald Coase. The paper was The Nature of the Firm.\ More than fifty years later, it won him the Nobel Prize in economics. His answer fits in one sentence: "The main reason why it is profitable to establish a firm would seem to be that there is a cost of using the price mechanism." Honestly, I only quoted that to sound smart. In plain English:\ **Going outside to find someone, haggle, and sign a contract for every single task is too expensive.**\ **Better to bring people together and manage them on a long-term, fixed basis.** Coase also named the most obvious of those costs:\ **just figuring out what the price should be takes effort.** When doing it inside is cheaper than buying it outside, the firm is born.\ Economists call this "transaction costs." Software works the same way. Take a fixed process, pay someone once to turn it into fixed logic.\ Running it from then on is far cheaper than doing it by hand every time. Contract review works like this. Inventory management works like this. So do hospital medical records.\ All standard processes. What did people do before there were systems? Copy by hand.\ Paper passed back and forth: hard to manage, even harder to find. So we got software.\ Pay a little, get a big jump in efficiency. ## 2. People want the result, not the process So why does software have so many buttons, windows, workflows, and clicks? Nobody wants to learn the hundreds of features in Photoshop. We're forced to.\ We just want the final image. Theodore Levitt of Harvard Business School,\ in The Marketing Imagination (1983), quoted an old line: "People don't want to buy a quarter-inch drill. They want a quarter-inch hole!" When a doctor enters a medical record, what do they actually want?\ For the record to land safely in the system.\ So that whenever the patient is discharged, transferred, or sent for a consult, it can be found. But to get that hole, they first have to learn the drill.\ Click this, click that. And hospital computers are usually painfully slow.\ Forget to save once, and it's another half hour before the record is in. Photo editing is the same.\ You don't learn all those techniques for the process.\ You learn them to turn a plain photo of a coffee cup\ into the ad poster your boss wanted. Now take email.\ In Outlook you open one message, then another.\ You scroll the mouse wheel for what feels like a mile.\ You find the PDF, download the attachment, open it, squint and read the whole thing.\ And finally tell your boss: here's the deal, blah blah... That one sentence is all he wanted. Design has a term for exactly this.\ Don Norman, author of The Design of Everyday Things,\ called the gap between "what I want" and "what the system forces me to do"\ the **gulf of execution**. Old software built a very hard-to-cross bridge over that gulf.\ People always wanted the result.\ Software made them learn to cross the bridge first, then handed over the result. Not anymore. ## 3. Agents arrived, and the process came right back Now we have agents, and MCP — a standard interface that lets AI read and write your data.\ As long as the AI can reach your data,\ you say one sentence and the result appears. In theory. In practice? First you have to understand what a large language model is.\ ChatGPT, Claude, Kimi, Qwen, DeepSeek — what's the difference?\ This one is MoE, that one isn't.\ This one has more parameters, that one tops the leaderboard.\ This one's great at code, that one's great at design. You finally pick one, sign up, and start using it. Turns out it doesn't quite listen. So you train it.\ Connect MCP so it can touch your data.\ Install skills so it remembers your routines.\ Add hooks to put guardrails on it.\ Then rewrite the prompt again and again. After a lot of tinkering, it finally runs smoothly.\ Almost no back-and-forth rework.\ One sentence a day, one daily report you're happy with. A long detour, all for the result. ## 4. Is this really what people want? Stop and think. Why should we have to understand whether you're MoE?\ Why should we have to know about parameter counts and reasoning effort?\ Why should we have to scour the internet for this skill and that skill? I just want to finish my work and go home early. Isn't this just the old software road again?\ Learn the process first to get the result.\ The process just changed from "clicking buttons" to "tuning AI."\ The gulf of execution didn't get any narrower. It's still a massive pain. Larry Tesler, a veteran Apple engineer, had a law: Every application has a certain amount of complexity that cannot be removed. The only question is who carries it: the user, or the developer? Today, that complexity sits on the user. Engineers who love DIY will adore all this.\ But for a technology to be usable by ordinary people, let alone used well,\ the engineers behind it have to carry the complexity on the user's behalf. ## 5. Technology always specializes first, then goes mainstream Every technology develops gradually. In Diffusion of Innovations, the sociologist Everett Rogers\ split the people who adopt a new technology into five groups:\ innovators, early adopters, early majority, late majority, laggards. The first group is only 2.5%. The people fiddling with skills, MCP, and hooks today are that 2.5%.\ They're willing to grind through the process to get the result. The other 97.5% won't.\ They'll wait for the day the process is hidden. Specialize first, then generalize, then go mainstream. Photography went this way.\ In the early days, you mixed your own chemicals\ and crawled into a darkroom to develop the film.\ In 1888, Kodak ran an ad:\ **You press the button, we do the rest.** Driving went this way.\ Early cars had to be started by getting out and turning a hand crank.\ If the crank kicked back, it could break your arm.\ In 1912, Cadillac installed an electric starter.\ Sit down, press, and it starts. Computers went this way.\ At first you had to memorize line after line of commands.\ Then came graphical interfaces: just click the mouse.\ Later still: just touch it with your finger. Every past technology went this way. AI will too. ## 6. The endgame: an agent market that sells only outcomes Once the process is hidden, how does it actually work? Here's an example. You want to sue a company.\ Today's way: open Kimi, install a skill, connect Tianyancha (a Chinese company-records service),\ just to figure out the other side's legal representative, ownership structure, and pending lawsuits before you file. Then you install a few legal skills and connect a few legal MCP servers,\ just to work out how to respond when the defendant starts weaseling. The future way:\ go to the market and find an investigation agent and a legal agent. You hand over only three things:\ **instructions, materials, context.** They use their own industry experience and deliver the result directly.\ Priced by outcome, paid by outcome. You don't need to know how a large language model differs from an agent.\ You don't need to understand MCP or skills,\ or know anything about MoE, parameter counts, or reasoning effort.\ You don't need to understand anything. None of it is any of my damn business — I just want the result. This has already started. Intercom, which builds customer-service agents,\ charges for its AI support agent per "resolved conversation": $0.99 each.\ Not resolved, no charge. Bret Taylor, co-founder of another agent company, Sierra, puts it even more bluntly:\ the whole market is moving toward paying for outcomes. In The Wealth of Nations, Adam Smith titled one chapter:\ **The division of labour is limited by the extent of the market.** The bigger the market, the finer the division of labor.\ When anyone can buy a result with one sentence,\ the market becomes big enough\ to support an agent that only reviews contracts, one that only does due diligence, one that only handles medical records. People pay only for value and outcomes.\ Never for process. ## Closing In 1911, the philosopher Alfred North Whitehead wrote: "Civilization advances by extending the number of important operations which we can perform without thinking about them." --- **References** Ronald Coase, The Nature of the Firm, Economica, 1937 Theodore Levitt, The Marketing Imagination, 1983 Don Norman, The Design of Everyday Things Larry Tesler, Law of Conservation of Complexity (c. 1984) Everett Rogers, Diffusion of Innovations, 1962 Adam Smith, The Wealth of Nations, 1776, Book I, Chapter III Intercom Fin pricing page: intercom.com/pricing Alfred North Whitehead, An Introduction to Mathematics, 1911 Originally published in Chinese on my WeChat account 罗里戴行思录: [read the original](https://mp.weixin.qq.com/s/owQcWuhFxdkOID1PtjtJWQ)