The delegation tree
The delegation tree decides which steps of a job a person keeps and which an agent takes. It is used one step at a time, never on a whole job at once.
Where you see it
Open any agent's agent.md and look at its job. Every step has a label: you, drafts, owns, or not yet. Those labels came from this tree.
What happens, step by step
The core rule is one sentence. An agent gets a step only when someone can check the result faster than they could do it themselves.
For each step, walk the tree in order:
- Done fewer than 3 times by hand? It stays the person's. One time is a data point, not a pattern.
- Can a check run without a person? A test, a comparison, a number. If a check cannot be built in about an hour, the person does it, or hands over only the mechanical part.
- Does it reach a real person, cost money, or is it hard to undo? The agent may draft. Writing to someone who is not on the agent's allowed list always needs the owner's yes, at any rung. A Proven agent may send to people on its list without a new tap. Spending follows the permission table. See The trust ramp.
- Will it happen fewer than 5 times, ever? The person does it. It is not worth writing up.
- Otherwise, the person does it 3 to 5 times to build the pattern, writes it down as a brief, and starts the tree again at step 2.
The weakest-link score. For a step that makes it through, score four things from 1 to 3:
| Factor | The question |
|---|---|
| Worth it | Does it happen often enough to hand over? |
| Agent can do it | Can the agent really do it well? |
| Person can check it | Can the owner check the result quickly? |
| Safe if wrong | Is a mistake cheap and easy to fix? |
The lowest score wins. It does not average. A step that scores 3, 3, 3, and 1 is a 1.
- Lowest is 1: label you.
- Lowest is 2: label drafts.
- Lowest is 3: label owns.
The four labels:
| Label | What it means |
|---|---|
| you | The person does it. The agent only hands over research and facts. |
| drafts | The agent drafts. The person approves. |
| owns | The agent does it and checks its own work. |
| not yet | Same as "you" for now, because it has been done fewer than 3 times by hand. Flagged as a future candidate. |
What powers it
| Part | What it does |
|---|---|
| The delegation tree | The ordered test above. A written rule, not code. |
| The weakest-link score | Four factors, 1 to 3 each. The lowest sets the label. |
agent.md | Where each step's label is written down. |
add-resident-agent | The hiring skill. It walks the owner through the tree for each step. |
Why it works this way
If the owner has done something once, there is no pattern yet. An agent cannot copy a pattern that does not exist.
The tree was written after a real mistake. In the reference install, Ada, the workshop agent, was handed the whole workshop process. The owner had only run a workshop once.
The logistics went fine: picking a date, checking the room, booking it, waiting on the local library to confirm. Those steps have one right answer and are easy to check.
The creative steps did not go fine. The curriculum, the invitation emails, the promotion, and the follow-ups "burned weeks of chat and never got good." There was no proven version for Ada to copy, and no quick way to check her work.
So now logistics are labeled owns. The creative steps are you or not yet, until the owner has done each one by hand enough times to know what good looks like.
The weakest-link rule exists for the same reason. A step can be frequent, easy, and quick to check, and still be a disaster if it goes wrong. Averaging would hide that. Taking the lowest score does not.
Connected to
- Agents vs routines: the first question, before any steps are sorted.
- The trust ramp: what an agent may do on its own at each rung.
- An agent's files: where the labels are written.
- Example agents: Ada's steps and labels.