Building an agent from parts

A new agent is not written from scratch. It is one or two work shapes, plus a shelf of shared parts every agent already has, tied together by a contract the owner agrees to in a short interview.

At a glance

Where you see it

A setup conversation with Claude. The result is a new folder for the agent and a new row on the Office.

What powers it

The add-resident-agent skill, the archetypes folders, the interview rows in the logbook, the plan door, the shared modules.

When it runs

Once per new agent. The owner can change the contract any time after.

Hiring a new agent

Shapes of the job

Filler

80%

Brings: a 5-step plan, the fill-rate number, when to wake.

Chaser

20%

Brings: follow-up steps, touch limits, its guardrails.

Shared parts every agent gets

Wake checklist wallRoom wallDecision captureEmail intakeDeleted folderSend guardThe Office row

→

Ada new

agent.md · the contract

The result

15 people in the room at every monthly baking workshop.

The number

Came to the workshop: 9 of 15.

Never

Promise a discount. Post in public without a yes.

Steps and who does them

Book the room · ownsSign-up page · draftsClass outline · you

Example screen with made-up data.

Where you see it

The owner sees a conversation, then a new agent on the Office with a face, a goal, a number, and a plan already approved. Behind it is a folder in the vault (the folder of plain text files that holds everything) named after the agent. Its agent.md is the contract. The owner can edit it directly.

What happens, step by step

Say the bakery owner says: "I want something that keeps my workshops full."

1. First, the four-kinds test. Is this started by a clock, an event, a person asking, or the agent's own judgment? Only the last one becomes an agent. A fixed list of steps becomes a cheaper routine instead.
2. Research first. Before any contract, Claude looks up how others do this job and writes what is worth taking and what to avoid.
3. Name the shapes. One or two, with a share of time. Here: Filler, mostly, with some Chaser.
4. The interview. The questions are rows in the logbook, asked one at a time: the result, the number, what it may never do, its tools and people, how often it wakes, how the job splits, and who does each step. Then any extra questions the shapes add.
5. Write the folder. agent.md (the contract), memory.md (what it learns), scoreboard.md (its number over time), craft.md (what makes a good one), a face, and colors for the office picture.
6. Set the first wake before anything else. This stops the agent from waking up before its plan exists.
7. Fill the plan. The shapes' starter steps and numbers go in through the plan door, named in the owner's words, each with its who-does label.
8. Sign. The owner signs the contract, and that one act approves every starter step. A budget, if any, is typed on the Office's Settings tab.

The shared parts

These are built once and every agent gets them. The agent does not have to know how they work.

Part	What it does for every agent
The wake wall	The agent cannot do anything until it has read its checklist: contract, plan, memory, recent history, inbox, and what is new on the forum.
The room wall	The same wall when the owner opens a chat with the agent.
Decision capture	A decision made in a chat is written to the plan, the inbox, or memory the moment it lands.
Email intake	The agent never reads a raw email. A tool-less model fills a form about it, and the agent gets the form.
The deleted folder	The agent never truly deletes. It moves things aside with a reason.
The send guard	Every message to a real person goes through one checked door.
The Office row	Every agent shows up on the page the same way, with no extra work.

What powers it

Part	What it does
add-resident-agent skill	The step-by-step setup Claude follows.
logbook.py interview show	Reads the interview questions, which are editable rows.
Resources/Agents/archetypes/	The 17 shapes: starter plans, numbers, guardrails, wake rhythms.
Resources/Agents/modules/	The shared parts, each with a plain page and a builder's page.
logbook.py plan	Where the starter steps and numbers are written, and where the plan is approved.

Why it works this way

Why the first wake is set first. One new agent was copied onto the box and woke up seven seconds later. It had no checklist and no steps yet. It spent 25 turns and $2.24 before anyone noticed. It did no harm, but the order changed: set the wake time first, then build the rest.

Why the interview is short. One pass through the questions, then stop. A long, open chat produces a vague contract, not a better one.

Why one signature. The owner read the steps in the interview and signed a contract that describes them. Asking the same yes again for each step was the wrong shape. Only steps the agent adds later need a new stamp.

Why parts, not code. A new kind of job should be set up, never coded. If a new agent needed new engine code, the design would be wrong.

Connected to

• The 17 kinds of work: the shapes a job is built from.
• The agent's files: what each file in the folder does.
• The trust ramp: what a new agent may do alone, and how that grows.
• The delegation tree: how each step gets its who-does label.
• Agents vs routines: the four-kinds test in full.