How to manage something brilliant and fast — and forgetful.
craftagent is where I write down what working with AI agents actually taught me: the principles, the architecture, the moves that held up — earned from my own faceplants.
Never handed work to an AI? Start here.
Familiar desk-pains — and the narrow doorway to your first handoff without freezing at the blank field.
Four gates on the life of one task.
An agent is a forgetful genius junior. Place a human-control gate at each spot it tends to fall — tap to see the prompt.
Twelve clusters of managing an AI agent.
Steering the agent
Make it understand the task before it acts — through command structure and forcing functions, not luck.
4 piecesEnter 02Context & memory
Memory is a finite resource: what to keep, what to drop, when to let a script do the work instead of the AI.
4 piecesEnter 03Orchestrating many agents
When one agent isn't enough: split roles, hand off work, gather results without chaos.
4 piecesEnter 04Shipping with AI
Four gates — plan, checkpoint, impact, verify — so an agent ships real work without breaking things.
5 piecesEnter 05Production in the wild
Faceplants from the field — retold with names stripped, so you can dodge them first.
8 piecesEnter 06Economics & tradeoffs
Tenfold speed is the sticker price — what you take home depends on the cost of rework, checking, and trust.
5 piecesEnter 07How you change
You do not learn how to work with an agent once and move on — you pass through four stages, and the dangerous one is not the beginning.
4 piecesEnter 08Tools & setup
Your setup determines what the agent can do — and CLAUDE.md is not documentation, it is the operating law of your agent.
4 piecesEnter 09Autonomy — how far to let go
"Can I trust it?" is the wrong question. The right one: for this task, with this reversibility, with these stakes — where does the gate go?
4 piecesEnter 10Measure & improve
Without a number there is no improvement — only a shifting feeling. Three metrics that can actually tell the difference.
5 piecesEnter 11Security & trust
When an agent touches a real system, the question shifts from "is it correct" to "what can it reach, and what leaves your hands".
4 piecesEnter 12Working as a team
Rules in one head don't multiply to a team — when many people plus many agents share one thing, the shared rules must be written and shared.
7 piecesEnterAn agent is not a machine you command. It is a genius junior with amnesia — you manage it, you do not let it run loose.
01 / Manage, don't release
A brilliant junior who types ten times faster than you, fearless — but wakes up having forgotten yesterday. So you supervise. And supervising means placing a guard exactly where it tends to fall.
02 / The real cost
It goes rogue once, you lose trust, then you re-read every line. That tenfold speed you just bought — gone. That is the real bill.
03 / Scaffolding beats model
The stronger the model, the more the control scaffolding matters. You don't need a smarter agent. You need to become a better manager.
Start from “Shipping with AI”.
Four control gates that let an agent ship real work
Use this when you want the agent to work more independently without quietly drifting off and calling it done