Nothing matches.
Not a loose list of pieces — a terrain. Every piece falls into exactly one cluster; twelve clusters stack into three tiers, from foundations to the field. Scroll to walk the whole map.
Make it understand the right task before it acts — through command structure, forcing functions, and a clarify-first habit, not luck.
Memory is a finite resource: what to keep, what to drop, where to persist it — and when to let a script do the work instead of burning tokens on the AI.
When one agent isn't enough: split roles, hand off work, gather results — orchestration patterns so many agents run without stepping on each other.
Four human-control gates — plan · checkpoint · impact · verify — placed at the four moments an agent is most likely to fall on the life of one task.
Faceplants from the field — retold with every name and specific detail stripped, keeping only the general trap so you can dodge it first.
The sticker speed and the take-home speed are two different numbers. This cluster measures the take-home one: when an agent actually pays off, when it loses money net, and what not to hand it.
The agent does not change. You change — through four predictable stages. This cluster maps them and lingers at stage three: where you are skilled enough to trust but not yet wise enough to know the limit of that trust.
Most people write CLAUDE.md like a README and wonder why they still have to repeat instructions every session. This cluster draws the line between describing and forcing function — and what a setup actually needs to work on its own.
Autonomy is not a switch. It is a dial — and the right position depends on three dimensions of the task, not on how you feel about the agent. This cluster gives you the framework to let go in the right places instead of letting go more or less.
"The agent has felt solid lately" is a metric of your memory, not of reality. This cluster gives you three numbers — catch rate, rework rate, autonomy ceiling — and explains why not measuring means not improving, only getting more comfortable.
Once an agent holds the keys to a real system — running commands, reading data, sending things — the worry is no longer correctness but what it can reach and what leaves your control. This cluster maps the five faces of risk in granting access, and the rule: narrow what it touches, record what it does.
The whole garden assumes one person. But when the whole team has agents on the same thing, conventions diverge, ownership blurs, craft doesn't transfer. This cluster is about pulling the rules out of each head into a shared place every agent reads.