What an agent may do without a person
Every agent we build has a level. The level says what it may do on its own and what waits for a person. We agree the level for each agent with the company before it runs.
Three levels
Read only
Reads and summarises. Nothing leaves your company.
Prepare and wait The default
Prepares each action. A named person approves it before it goes.
Act within limits
Acts alone inside limits you set, such as known customers and an amount cap.
- Read only. The agent reads and summarises: a report for the owner, an answer drafted for staff. Nothing leaves the company.
- Prepare and wait. The agent prepares an action and waits. A named person approves each one: a message to a customer, a quote, an invoice, a payment reminder. This is the default.
- Act within limits. The agent acts alone inside narrow limits the company sets in advance, such as known customers and an amount cap. Anything outside them waits for a person.
Why prepare and wait is the default
Most agents we build read messages from people outside the company, see the company's records, and can send messages or create invoices. Meta's security team argues that an agent with all three should not run without a person supervising it [1]. So by default, a person approves before anything leaves the company or any money moves.
One approval for each message
When many messages wait, they are sorted by risk, so the person deciding sees the ones that matter first. Each message to a customer still gets its own approval. There is no button that approves a whole list.
Where the levels come from
The levels follow the approach in AWS's scoping matrix for agentic AI security, which sorts agents by how much they may do on their own [2].
Sources
- Meta, "Agents Rule of Two: A Practical Approach to AI Agent Security", 31 Oct 2025. ai.meta.com
- AWS, "Agentic AI Security Scoping Matrix". aws.amazon.com