High Intent Labs
Basics · AI Agent Builder Academy · 23 Aug 2026
ENDE
Plain English

Harness, skill, loop, graph. What they mean.

These words sound big. They describe small, concrete things. Here every word sits next to the part it names, with a real example from a setup that runs every day: mine.

Instructions
21 skills
10 of them written by me
Roles
17 subagents
each with its own context
Automation
19 timers
plus one hook at session end
The flow

Three ways in, one path through

Any one of the three inputs starts it. After that it always runs the same way: load the context, work in a loop, hand off the big parts, deliver.

How the flow runs from input to result Three inputs, each on its own, can start the flow: you typing, an email arriving, or the clock. They merge into one path. Step one loads the context. Step two is a loop of think, use a tool, read the result, which repeats until the job is done and hands large parts to helpers. Step three delivers a note, a live page or a draft. INPUT · ANY ONE OF THESE STARTS IT You you type a request or An email arrives in the inbox or The clock 7:10 every morning 1 · IT LOADS YOUR CONTEXT Your rules, its memory, one skill loaded before it does anything 2 · IT WORKS IN A LOOP Think what is needed now Use a tool mail, files, the web Read the result did it work? it repeats until the job is done For big jobs it hands work to helpers they work in parallel each gives one result 3 · YOU GET A RESULT A note written into memory A live page updated online A draft email, deck, offer Navy = context · Orange = work · White = in and out
The loop is where the time goes. It keeps going until the job is done, and for anything large it hands parts to helpers and takes their results back. Nothing moves on to step three until the loop is finished.
The decoder

Eight words, one example each

What it is, in two sentences. Then what it looks like in real work.

Harness

Everything around the model: what it knows, what it may touch, what rules it follows, when it starts. The model is the engine. The harness is the car.

ExampleA better model comes out. I switch it on and everything keeps working, because none of my setup lives inside the model.

Context

The information the model can see right now. It is limited, so the job is to give it the right part, not all of it. More context does not mean better answers.

ExampleMy rules are always loaded. My 130 notes are not. Only their table of contents is, and the matching note gets pulled when the topic comes up.

Skill

A written instruction that gets loaded only when it is needed. Until then the system just knows the title. A skill is know-how, not a tool.

Example"Sort receipts for the accountant." Forty lines: where files live, how to rename them, what the tax office needs. Written once, followed the same way every time.

Subagent

A second worker with its own memory, its own job and its own tools. It works on the side and hands back one result.

Example"Research this company." It reads thirty pages and returns half a page. The thirty pages never enter my chat, so my chat stays sharp.

Graph

A fixed sequence of steps, written in code. The code decides the order, not the model. That is the whole difference to "the agent will figure it out".

ExampleEvery larger request runs the same six steps: plan, work, test, review, compliance check, result. Same order on a good day and a bad day.

Loop

Three different things carry this name: the basic rhythm of think and act, the way back when something is not good enough, and repeating on a timer.

ExampleThe compliance check says no. The work goes back to the role that caused it, not to the beginning. After three rounds it stops and comes to me.

Hook

A trigger tied to an event, with no model involved. It runs every time, exactly the same way. The least glamorous and most reliable part of the whole setup.

ExampleWhen a conversation ends, it is written into my notes automatically. No decision, no prompt. That is why I have never lost one.

MCP

One standard plug for outside services. Build the plug once and every service speaks the same language to your agent.

ExampleMail, calendar, drive, Notion, GitHub and the business bank account are all connected the same way. Adding the next one takes minutes, not a project.
The loop that matters

Not repeat until it is done. Improve until it is good.

Inside the loop above, a big request runs these six steps in order. Each step only starts when the one before it passed. If a check fails, the work goes back to step 2, not to the start.

Six steps with a way back Plan, then fan-out, then test, then a check of the whole picture, then a privacy and legal gate, then one result. A dashed line runs from the gate back to the fan-out, limited to three rounds. 1 · Plan split it, set the bar 2 · Fan-out roles work in parallel 3 · Test pass or fail 4 · Whole picture does it answer it 5 · Gate privacy and legal 6 · Result one answer, one voice back, max 3 rounds Orange = work · Navy = check · White = output
Three places can send work back: a failed test, a result that misses the question, and a red privacy check. After three rounds it stops and comes to me with a clear note, instead of looping forever.

This is the difference between an impressive demo and a system you trust with real work. A demo runs once. A system has a place where someone is allowed to say no, and a defined way back.

Order

Build it in this order

The common mistake is starting at the end. Automate first and you automate the wrong work.

Context first. Who you are, how you work, what has already been decided. Without it, no tool helps.
Then tools. Only once the system knows your context should it touch your systems.
Then skills. Anything you have explained more than twice gets written down instead of repeated.
Then roles. A separate worker with its own context pays off once the task is clear.
Then the fixed sequence. Wire down what has to run the same way every time.
Loops last. Only repeat automatically what has already proven itself by hand.
Honestly

Three words you will not find in my setup

Vector search

My notes are searched like a folder, not by embeddings. At this size that is faster, cheaper, and I can always see why something was found.

Evals

No automated scoring. Quality comes from the way back and the gate. Judgement, not a number on a dashboard.

More graphs

There is exactly one. Everything else is a skill plus a role. A second one gets built when a process actually demands it.

A good setup is not the one with the most parts. It is the one with the fewest parts you actually need.

Nicolas Meibohm
Nicolas Meibohm
AI Agent Builder Academy

I build startups and innovation programmes, and I run the setup shown above every day. In the academy you build your own in four weeks, on your real work, with live guidance.

Build this for your company

Four weeks, remote, live. At the end your setup runs, not mine.

Join the academy Book 30 minutes
highintentlabs.com · questions: nico@highintentlabs.com
Nicolas Meibohm · High Intent Labs
Plain-English guide to agent setups · 23 Aug 2026 · all numbers from the running setup