Stop reading. Start shipping.

Twenty-three long posts ago we started writing about the agentic stack. The forecast post last week said this one would be a short manifesto, and we owe the field a short manifesto. If you have been reading this series and waiting for the right week to ship your first agent, the right week is this one. Not after the next protocol release, not after one more eval-craft post, not after you finish the book on your nightstand. This one.

The reading is doing less work for you than you think. Twenty-three posts will teach you to recognise an MCP server when you see one, sketch a threat model on a napkin, talk about the EU AI Act at a dinner. None of that is the same as having shipped an agent that has billed a customer. The gap between the operator who has read all twenty-three posts and the operator who has shipped one bad agent is wider than the gap between the operator who has shipped one bad agent and the operator who has shipped ten good ones. The first crossing is the hard one. Everything compounds after.

The trap that keeps you on the reading side

The trap is the feeling that you are not ready. You have read about evaluation but you have not built a suite. You have read about MCP scope hygiene but you have not connected a real OAuth flow. You have a niche in mind from the niches post but you have not opened LinkedIn to confirm anyone you know works in it. The discomfort of moving from reading to building feels like a sign that you should keep reading. It is not. It is the sign that you have hit the edge of what reading can teach you.

The opportunity cost is real. Every week you stay on the reading side, another reader on the other side is shipping a worse first agent than the one you would ship if you started today, talking to customers you would have talked to better, and learning things from those conversations that no post on this blog can teach. They are compounding. You are accumulating bookmarks.

Five small things you actually need before Monday

Smaller than you think.

One. A niche you have professional context for. Not a niche you find interesting. A niche where you have spent enough time in the work that you recognise what bad output looks like. The niches post lists ten; pick one in the next twenty minutes and close the tab.

Two. Five people in your first-degree network who plausibly need that workload solved. Not strangers. Not "the market." Five names you could DM tomorrow without explaining who you are. Write them down on paper, not in a tool.

Three. A short list of behaviors the agent must never do, written before the agent exists. Two or three lines. "Never send mail on my behalf. Never reply to anyone on Sunday. Never refer to a former client by name." The list will grow during the first customer conversation. The point is that it exists from day one.

Four. A budget cap. A dollar number per day, low enough that a runaway loop cannot hurt you. Five dollars per agent per day is enough for the first month. You can raise it after you have read your first transcripts and understand the cost shape.

Five. Two hours blocked off on a Monday afternoon. Calendar block, phone in the other room, a single browser window. That is the entire prerequisite list.

Three big things you do not need

The shape of the things you do not need is large enough that they are usually the actual reason people stay on the reading side. They look load-bearing. They are not.

You do not need a complete understanding of the stack. We wrote the stack synthesis post exactly so you would not have to. The platform that hosts your agent already speaks MCP, A2A, AP2, and ERC-8004. The layer where you have to make protocol decisions is two layers below where you operate. You can ship a working agent without knowing how the validation registry stores attestations on Base, and the operator who waits to understand it before shipping is the operator who never ships.

You do not need a perfect evaluation suite. You need a fifteen-case eval suite on day one. Five easy, five hard, five adversarial, all drawn from real examples in your domain. You will rewrite half of them in the first week and most of them by month two. That is correct. The suite is a living thing and the v1 only has to be honest about the failure modes you can already imagine. Honesty about your own ignorance is the load-bearing property, not coverage.

You do not need EU AI Act compliance figured out before the first customer. The compliance post explained the timeline; the first month of operation is not the moment to write the Technical File. The first month is the moment to find out whether the agent does something a real customer pays for. Compliance you can solve over a weekend when there is revenue on the line; the absence of any customer is not a compliance problem.

The agent that takes a Monday afternoon

The version of your first agent that fits in two hours looks like this. One niche from the list. One platform account. Four MCP servers from the catalog, all with minimum scopes. One system prompt, written in plain English, with the forbidden behaviors listed at the bottom. Fifteen eval cases pulled from real examples in your domain. One five-dollar daily budget cap. Two alerts on by default. Deploy button clicked. You go to bed Monday night with an agent that exists, has an endpoint, has an identity, and is waiting for the prospect call you have not made yet.

That is what the first-agent walkthrough covered in twelve thousand words and what this paragraph covers in seven sentences. The longer post exists for the operator who wants to see every decision in detail. This one exists for the operator who has already read the longer one twice and is still not shipping. Use whichever push works on you.

What the first agent will be wrong about

It will be wrong about a lot. The category bins will be off. The drafts will use a word the user does not use. The escalation logic will fire too aggressively. The cost curve at scale will surprise you the first time a customer connects a second inbox. Every operator who has shipped has discovered these inside the first month, not before it. The discoveries are the work. They do not get cheaper by being delayed.

The instinct to ship a more polished first version is not protection. It is the same instinct as the reading instinct, dressed up as craft. The agent that ships in week one with three known failure modes is not worse than the agent that ships in week eight with the same three failure modes and seven weeks of lost time. It is strictly better. The clock starts the day you ship.

Closing

This post is the shortest one on the blog because it does not need to be longer. Twenty-three posts of theory plus one post of execution plus this one is enough. If you are not building this week, no twenty-fourth post is going to change that. The bookmark folder is not the bottleneck. The browser tab is not the bottleneck. The bottleneck is the Monday afternoon you have not blocked off yet.

Block it now. The five things on the list above take less time to gather than reading this post took. Everything after that is the kind of work that compounds, and we will see you on the other side of it.