Skip to content

What It Does Not Do

Reference

A capability list tells you what something can do. The exclusions tell you what it is.

Each of these was considered and refused. Several were refused after trying them first.

Not thisWhy not
Send mail as a member of the householdAttribution and reversibility. It drafts; I send
Spend moneyFast, repeatable, and hard to reverse
Post publicly or on social mediaAnything public is permanent in a way household work is not
Decide anything medical, legal or financialIt gathers and organises. A person decides, with professional advice
Be the source of truthHousehold knowledge must live in files I can read without it
Install or configure its own softwareSelf-modification destroys your ability to reason about it
Bypass my approvalConvenience here is how the project fails

The instinct is that a more capable agent is a better one. In a household, the opposite is closer to true.

Capability you cannot audit is capability you cannot trust. Every additional power is another thing to reason about, and reasoning has a budget. An agent whose limits you can hold in your head is one you will actually delegate to.

The exclusions are the product. Anyone can connect an agent to everything. The value is the judgement about what to leave out, and that is exactly what a stranger cannot copy from a feature list.

Refusals build trust faster than successes. The first time mine declined something it should have declined, the household’s confidence went up sharply. That only happens if you have written the refusals down.

If you cannot state, in one sentence, what your agent will never do — you do not yet know what you have built.

Not every exclusion was decided in advance. Both of these were retreats.

A capability to act more broadly on my behalf. It worked. It was also impossible to explain afterwards which actions were the agent’s and which were mine. I narrowed it back.

Automated document production that quietly accumulated copies. It ran perfectly for months. It produced an enormous number of files, answered no questions, and made it harder to find the documents that mattered. I kept one current copy and deleted the rest.

Worth naming both, because the useful lesson is not “I thought carefully”. It is that some capabilities only reveal themselves as mistakes once they are running.

The four questions I ask before granting anything:

  1. If this went wrong at 3am, would I find out before the morning? If no, it waits.
  2. Is it reversible? If no, a person stays on the path.
  3. Can I take it away cleanly? If withdrawing it breaks other things, it was not scoped properly.
  4. Could I explain what it did, to someone else, afterwards? If not, I have lost the thread.

Three yes answers and a stated purpose is my bar.