Skip to content

Notes · 2026-10-19

Agents propose, people dispose

Why every write in my agent systems is a proposal that waits for a human.

Every agent system I build has the same rule: an agent can read, and it can propose a write, but nothing is written until a person approves it.

It shows up everywhere:

  • In production work, specialists act only through scoped tools, and a human approves before anything changes.
  • In Tally Stack, the accounting assistant turns plain English into a proposed transaction; it is posted only when someone presses Confirm. The model never touches the database.
  • On this site, the demos follow the same rule. The Run Desk on Sheet 04 pauses on "approval required", and nothing is committed without it.

The rule costs a click. In return you get an audit trail of who approved what, a natural place for review, and a system you can trust enough to give real permissions to.