Skip to content

Adding More Tools

Step 4 · Its Tools

After calendars and email, everything else is optional. This page is the rule I use for deciding, so you do not end up with an agent that can do everything and cannot be reasoned about.

One tool at a time, only when a real task has been blocked without it, with a stated purpose and a named owner who wanted it.

Capability is easy to add and hard to take away. Every tool is a door, and an agent with fewer doors is one you can hold in your head.

Before granting anything: if this went wrong at 3am while I was asleep, would I find out before the morning? If the answer is no, the capability waits.

ToolWhat it unlocksAccess
A task managerWork that spans days and peopleRead and write, its own project
A web browserBooking pages, availability, forms that need a sessionRead, and act only where reversible
Web searchLooking things up without an accountRead only
A shared notes and knowledge storeHousehold knowledge as readable filesRead and write
A smart-home systemWatching the house, and reporting the unusualRead mostly, and only the sensors that matter
A grocery serviceBuilding the weekly basketAdds to a cart. Never checks out
Media automationFinding and retrieving things I asked forRead and write within its own library
Finance dataLive prices for a daily summaryRead only, and no trading access

Notice each row is narrow. The grocery row adds to a cart and cannot check out. The finance row reads prices and cannot transact. That narrowing is the work.

RefusedWhy
Sending email as a personAttribution and reversibility
Anything that spends moneyFast, repeatable, hard to reverse
Posting publicly or on social mediaAnything public is permanent in a way household work is not
Acting as the source of truthKnowledge must live somewhere a person can read without the agent
Administrative control of its own machineSelf-modification destroys your ability to reason about it
Automatic deletion of anythingDeletion is the one thing you cannot review afterwards

Regret is asymmetric here. I have never wished the agent could do more. I have twice wished it could do less, and removing a capability after it is in use is much harder than never adding it.

One agent with everything is convenient and impossible to audit. What I do instead is split by lane, with a separate agent per area of the household — one for general household logistics, one for family, one for health, one for business.

Each has its own memory, its own tools, and its own boundaries. None reaches into another’s systems.

  • One lane means one blast radius. A mistake in the health lane cannot reach my business mail.
  • Privacy becomes structural, not promised. I can tell the household exactly what each one can see.
  • It stays explainable. A single all-powerful agent is impossible to hold in your head.

This is more setup. It is worth it the first time you want to give one lane access you would not give another.

  • You can list every tool the agent has, and the job each one serves
  • Each tool has its own credential, and you can revoke one without breaking the others
  • Nothing it holds is irreversible
  • There is at least one thing it plainly cannot do, and the household knows what it is