The three guides here cover one task's life with an AI agent, from its first level to the weekly read of what it did. The first sets the level the task holds, the second covers the approval card and the stop switch around it, and the third covers the receipt each action leaves and the weekly read of those receipts.
The record of approvals, edits and rejections on a task decides when its level moves up, and one serious error moves it down. On August 1, 2012, Knight Capital's order router sent orders nobody intended for about 45 minutes, and the Securities and Exchange Commission later found the trading firm had no procedure for halting it while it ran (Release 34-70694); the levels guide tells the case.
What the guides cover
- autonomy levels, and the rule for moving a task up or back down
- approval cards, batches and the stop switch
- receipts, how long they are kept and the weekly read
- the cost of each completed action, in money and in a person's minutes
What they leave to other sources
Product how-tos are left to each vendor's documentation, which changes with every release. Attacks on agents are a security subject, and what an outreach agent says to a prospect is a sales subject. A guide that touches either says so at that point.
One task through all three
Each guide ends at the matching part of the agent readiness worksheet. One repeating task carried through all three comes out with a starting level, an approval card, a stop switch and a receipt design on paper. The worksheet also carries the weekly arithmetic: minutes in the approval queue, minutes for the weekly read and cost per completed action, set against a time budget written down in advance.
Reading path
- Levels of AI agent autonomy and how to raise one
Four working levels for what an AI agent may do on its own, and the written evidence that moves a task up a level or back down.
- Approvals and a stop switch for AI agents
An approval queue that fits in a few minutes a day, and a stop that halts an AI agent from a phone in about thirty seconds.
- What an AI agent audit trail should record
The fields an AI agent's receipt holds for each action, and a weekly read of those receipts that fits a fixed block of time.
Worksheets for this guide
- Agent readiness worksheet for one workflow
A printable worksheet that walks one repeating task from its first autonomy level to a weekly check of cost and time.