How a project runs

A project in Company OS is a goal that an agent holds for weeks. It has a few numbers to hit, a list of steps, and a label on every step that says who does it. The owner watches it on one page and steps in only where a step is theirs.

At a glance

Where you see it

The Office. Each agent's row shows its number and next step. Tap it for the full plan (the Roadmap) and your own steps.

What powers it

The plan door in logbook.py, the logbook database on the box, resident-wake.py, office_page.html, inventory.js.

When it runs

The agent reads and updates its plan on every wake. The page reads the same rows every time it opens.

The Office

Since you were hereAda finished "Book the room". Priya added 4 cafes to the list.

Needs your decision1 question from Ada, and 1 plan step

Ada needs you

Fill the spring baking workshop at Maple Street Cafe

Came to the workshop9 of 15

Waiting on the cafe to confirm the room

back Tue 9:00 AM

Priya asleep

Find local shops that want a catering partner

Leads you accepted6 of 12

Research the next 10 shops

back tomorrow 8:00 AM

Example screen with made-up data.

What you get

• One goal per project, with numbers. Not "more sign-ups" but "15 people in the room, every month."
• A plan of plain steps. Each step is named the way a business owner would say it: "Book the room", "Send the reminder email".
• A name on every step. The agent does it, the agent drafts and you approve, or it is yours.
• Nothing is done until someone signed off. A step nobody approved cannot be marked done, by anyone.
• Steps that wait for each other. A step can wait on another step, even one in another agent's plan. It becomes ready by itself when that step is done.
• Work between your visits. The agent wakes, works the next ready step, and picks its own next wake.
• One page to watch it all. Your part shows as "Yours to do", and every decision comes to you as one plain question.

Where you see it

Everything lives on the Office page. Each agent has one row. The row shows the agent's goal, a bar for its main number ("Came to the workshop, 9 of 15"), and the step it is on now, with what it is waiting for. If a step is yours, a button under it says Yours to do and names the step.

Tap the row to open the agent's sheet: every number, the step it is on, and Your steps. On the Your agents tab, tap an agent to see its whole plan as a rail of steps. That view is called the Roadmap.

What happens, step by step

Say a bakery owner wants the monthly baking workshop full. An agent named Ada holds that job.

1. The goal is written down. When Ada is set up, the owner answers a short interview. The answers become her contract, agent.md (the file that says what she is for and what she may never do).
2. The numbers go in. Each number is a key result (a count that shows the goal is working), like "Came to the workshop: 9 of 15." Each one says where the count came from.
3. The steps go in. "Pick the date", "Book the room", "Draft the sign-up page", "Send the reminder email". Each step has a state, a label for who does it, and a check that says when it counts as done.
4. The owner approves the plan. Signing the contract stamps every step at once. A step the agent adds later needs its own stamp.
5. Ada wakes on her own. She reads her plan first, then works the next step that is ready. If a step needs the owner, she asks one plain question and moves on.
6. She writes back. A finished step turns done. A step waiting on the cafe turns waiting, with the date she will nudge them.
7. She picks her next wake. If there is more she can do, she comes back in about an hour. If she is waiting, she sleeps until her own rule says to look again.
8. The owner steps in by answering a question, approving a step, or tapping Mark done on a step that was theirs.

What powers it

Part	What it does
agent.md	The contract. Holds the goal, the guardrails, and the steps with their labels.
key_results table	One row per number the project is judged on: target, current count, and where it came from.
milestones table	One row per step: its state, who does it, what it waits on, its checks, and who approved it.
logbook.py plan	The plan door (the one way in). show, upsert, set-state, note, kr-set, approve, signoff, next-wake.
resident-wake.py	Checks every 15 minutes which agents are due, and wakes them.
office_page.html, inventory.js	Draw the rows, the sheet, and the Roadmap from the same rows the agent writes.

Why it works this way

The plan used to be a file the agent edited. Nothing could check it. The owner could not change it from the page. So the plan became rows in the logbook (the database on the box), with one door in. The agent and the page both go through that door, and the door enforces the rules. That is how "nothing is done that nobody approved" can be a fact and not a promise.

The page shows steps, not just the final number, on purpose. The owner said: "the end goal is only interesting once a month; what I'm actually interested in is where it is in the progress of all the milestones."

The one idea

The agent does the work. The plan says who does each step. The owner only touches the steps marked as theirs, and the decisions an agent brings to them.

Connected to

• Plans and the Roadmap: the steps, their states, and the number bars.
• Steps, waiting, and sign-off: who does each step and when it may be marked done.
• Teams and handoffs: when one project needs more than one agent.
• How an agent wakes: what the agent does each time it wakes.
• The Office: the page where all of this shows.