Key Concepts
Scoped autonomy: Autonomy is a dial, not a switch. For each thing an agent can do, you choose whether it asks first, goes ahead on its own, or never does it, and you loosen the dial in steps.
Approval gate: A point where the agent pauses and a person approves before it continues. In Profound, this is a tool set to Ask.
Quick Reference
Step 1: Define what wrong looks like
- Write down what a bad result would be before the agent runs
- Examples: a claim without a source, a number that does not match the data, a message sent to the wrong group
- If you can describe a bad outcome, you can check for it
Step 2: Set a permission level for each tool
- Ask: the agent checks with you first
- Allow: the agent goes ahead on its own
- Block: the agent never does it
Defaults to start with
- Reading: usually safe to Allow
- Writing, sending, or publishing: start on Ask while you are watching the agent
- Riskier actions: keep on Block
Before the agent runs on its own
- An agent running on its own cannot answer an approval prompt
- Deliver through a low-risk step, such as a link to a doc
- Keep riskier actions on Block
- Move a delivery action to Allow only once you trust it
Step 3: Move up a level
- Start with the agent drafting and a person approving
- Wait until its work has been right long enough that you stopped changing it
- Loosen one permission
- Watch, then decide whether to keep it
- Repeat
Common failure to avoid
- Letting an agent publish directly because it has worked for weeks. One error that answer engines pick up shows the agent did exactly what it was allowed to do. Nobody had decided what that should be.
Notes
- The goal is not to withhold freedom from agents. It is to give the right amount at the right time.