Playbook
One person. Many agents. Run by values.
Values are fuzzy. We break them down, layer by layer, into goals and rules an agent can actually follow, then let a team of agents with clear jobs push the work forward and cover for each other. You only handle what only you can decide.
This is how the author runs things every day, and it keeps changing. It describes how we work today, not a product promise. Role names and examples have been made generic.
The map
One loop: values go in, interruptions go down.
- 1
ValuesWhat you want and what you need to protect. Different for everyone, and they change with life stages.
- 2
RulesFuzzy values written down as goals and limits an agent can follow: what it decides on its own, and what it must ask you.
- 3
AgentsA team of agents with clear jobs that cover for each other, working within machine, quota and time limits, and finishing whatever they can.
- 4
YouOnly what only you can decide, one question per card. This is the only place your attention goes.
Every answer you give and every mistake an agent makes goes back to step 2 as a new rule. So the longer you use it, the less it asks.
Two modes
Where your attention is decides which mode the system runs in.
Focus
You’re deep in one project, working with one agent.
- The system steps back to the minimum: just keep the cards right
- One question per card, only things only you can decide
- Every answer has someone responsible for picking it up
- In progress: if you already said “go” in the session, you won’t be asked again in the inbox
Juggling
Lots of agents running, many threads, and you’re not sure what to do next.
- Agents scan your projects against the values you wrote down
- They only suggest a next step when quota and your attention both have room
- Ranked by value, return and urgency; a few at a time
- In progress: each suggestion passes the “would this look dumb?” review first
Life of a task
From something you said in passing to a result back in front of you.
You mention an idea. An agent that collects ideas writes down your exact words and gives a judgment: worth it or not, how it relates to existing projects, the smallest first step, what it costs.
You say “go”. It becomes a one-page brief and an agent task. If it costs nothing and isn’t public, we’re making it so you aren’t asked again.
A dispatcher agent assigns it. Whatever agents can do goes to an agent; whatever needs you in person becomes a card.
An agent works on a branch. When it needs your call it sends one card with one question, and does everything else as far as it can.
A different agent reviews it. The one that writes code isn’t the one that reviews it, ideally not even the same model vendor.
Merge and ship. A dedicated role (also an agent) merges to main, deploys, and confirms it’s actually live.
Report back. One plain sentence with the result. If it isn’t finished, it says how far it got, what is left, and when.
Roles and what they prevent
Every role in this table was added after something went wrong.
A representative few. Each row is something that actually happened to the author.
| Answer pickup | You answered a card on your phone, but the session that asked had long closed, and nobody picked up the answer. Now there are four layers of fallback, and if none catches it you get a reminder within 30 minutes. |
|---|---|
| Relay | A session filled its context, wrote a handoff, and then sat waiting for someone to clear it, for seven hours. Now it writes the handoff and starts a fresh session to carry on by itself. |
| Follow-up timer | An agent said “I’ll remind you in 24 hours”, wrote it in a report, and nobody reminded anyone. Now every promise is registered on the spot and fires on time. |
| Patrol | Fixed code never shipped; a scheduled job quietly stopped. Now a patrol runs every few minutes, logs what it finds, and opens a fix branch. |
| Card check | One card crammed the question and both options into its title, which then got cut off. Now every card is checked as it’s sent; unclear ones get flagged or sent back, and an hourly spot-check scores the rest. |
| Incident log | The same mistake happened twice. Now serious mistakes are logged and reviewed on the spot, then turned into a rule or a mechanism. |
| Watchdog | The always-on machine restarted overnight for an update, its network didn’t come back, scheduled jobs failed all night, and nobody noticed until the morning report never arrived. Now there are self-checks and alerts, held until morning. |
The dumb check
When agents suggest where to go next, one dumb suggestion and you stop reading them.
The classic dumb move: ninety percent of the time you spend working with agents goes to one thing, so the agent concludes that thing matters most to you. But working with agents is a small slice of life. Spending a lot there only means that kind of work needs agents.
- No conclusions from a single angleTime and money spent are shown, never counted as evidence. A suggestion needs at least three independent kinds of evidence pointing the same way.
- Once you’ve said your piece, it’s settledIf you picked “not now”, it isn’t raised again unless the evidence changes.
- Always come with a change an agent can makeThe card is a yes/no; one tap moves it forward. “Let’s talk” doesn’t count.
- A review before it is sentAnother agent answers one question: would you find this dumb? If yes, it isn’t sent, and the reason goes into the rules.
The first time this ran, both drafts were stopped at review and nothing was sent. That’s exactly what it’s for.
For now it only covers one kind of suggestion: when the priorities you wrote down don’t match what you actually do. We’re extending it to all suggestions.
Safeguards
Agents finish whatever they can on their own. Only a few kinds of things reach you.
- Quiet at nightOvernight (the author uses 9 pm to 8 am), alerts and reminders wait until morning, except anything about money.
- A gate on spendingAnything above a small amount, or past a daily total, asks you first.
- You see it before the world doesAnything public, including this page, gets a preview for you before it goes live.
- Resources you can seeMachine memory and disk, and how much subscription quota is used, all have readings. When quota is tight, work waits in line; it’s never pushed onto you.
- Low-value work comes as one card (in progress)Things that could be done but aren’t really worth it are neither done on a whim nor asked one by one. Once a day they become a single “do these or not?” card.
Where to start
One inbox and one dispatcher. That’s enough.
The roles above grew over time, as one person used the system and ran into dozens of incidents. You don’t need them all on day one. Start by having every agent reach you through the same inbox, then add a dispatcher that takes over after you answer. Add a role when something goes wrong.
See pricing and join the waitlistConnection steps coming soon