Trecho da aula

What an AI Agent Harness Is Made Of: Instructions, Tools, Environment, State, Feedback Loops

Five harness engineering subsystems: instructions (AGENTS.md), tools, environment, state and feedback loops. Rollout order, ablation testing of each component's value at a fixed model, and the flow diagram between subsystems.

the five harness subsystems ablation measurement adoption curve

O que você aprenderá nesta aula

Break your harness down into its five subsystems and identify which one is missing.
Learn to measure the contribution of a specific component using an ablation experiment rather than by eye.
Draw up a prioritised implementation plan based on the returns curve.

Início da aula

'Just write a longer prompt'

The team spends a week writing a five-page AGENTS.md
The agent still won't run the tests by itself
Every session starts with fixing the environment
It looks like the model's fault — but it isn't

The team decides: since the agent keeps getting things wrong, everything needs describing in more detail — so they spend a week writing a five-page instruction document, listing libraries, tone of voice and the tech lead's favourite patterns. The agent reads it once, and still doesn't run the linter before committing — because it physically can't run the linter: it has no shell access with the right permissions. The problem isn't the length of the prompt. The problem is that instructions are only one of the five subsystems of a harness, and if the other four aren't in place, no amount of text will save you.

Myth: a harness = a prompt file

🤔 Myth / Common confusion: 'I've got a harness' = 'I've got a prompt file'
✅ In reality: a harness is five distinct subsystems, and instructions are just one of them
📌 Remember: a shelf full of recipes is not yet a kitchen

The phrase 'I've got a harness' almost always means 'I've got a prompt file' — an AGENTS.md or a CLAUDE.md of some kind. But that isn't a harness, any more than a fridge full of ingredients is a restaurant. You've got the ingredients, but there's no stove, no knives, no way to taste-check a dish, and no quality-control system. A proper harness is five working parts, and missing any one of them doesn't make things 'a bit worse' — it creates its own distinct, predictable class of failures that will repeat again and again.

A kitchen with five stations

A harness = five subsystems working together
A missing part isn't 'slightly worse' — it's a systematic failure
Next up: one station of this kitchen at a time

Take away just one station from a professional kitchen, and it either never opens or ends up feeding customers on a wing and a prayer. A kitchen has five stations: the recipe shelf, the knife station, the stove, the prep counter, and the pass where quality control happens. It's exactly the same with a harness: instructions, tools, environment, state and feedback only work together, and a weak station can't be made up for by a strong one.

A aula completa é para assinantes

Mais 19 slides na versão completa.

O vídeo, todos os slides, anotações, testes, tutor de IA e entrevista ficam disponíveis após entrar e assinar.

Entrar ou cadastrar-se Assinar

Conteúdo da aula

'Just write a longer prompt'
Myth: a harness = a prompt file
A kitchen with five stations
Stations 1–2: instructions and tools
Stations 3–4: environment and state
Station 5: feedback — the pass
Where to start: 5 → 1 → 4 → 3 → 2
❓ Quick check: which station is your weakest?
💡 Diagnosis: how to tell what's broken
Why different teams all follow the same path
🛠 What this looks like in real teams
How to measure a component's value
A zero drop ≠ 'the component is useless'
Ablation pitfalls: it's not the place to start
Constrain, don't micromanage
A map, not a textbook
Your harness rots just like code
Where an agent's work leaks when one link breaks
Read the diagram as a loop, not a list
30 minutes on the weakest station of your harness
Next up: your first AGENTS.md for a payments backend
A harness is a loop, not a file