Risk classes and autonomy levels
Every tool has a class from R0 to R4. An agent acts alone up to its level’s ceiling; above it, a person decides. R4 always waits.
Updated 1 min read
Two things decide whether an agent’s action runs by itself: the risk class of the tool, and the agent’s autonomy level.
Risk classes
| Class | Kind of action |
|---|---|
| R0 | Read, search, report. |
| R1 | Draft, note. |
| R2 | Reversible changes in your own systems: update a record, organise, message your own team. |
| R3 | Reach known people, publish, change settings or credentials. |
| R4 | Irreversible or outward: contact strangers, send in bulk, move money, delete, make a legal commitment. |
Autonomy levels
| Level | May do alone | Ceiling |
|---|---|---|
| L0, observe | Read and report. Everything else waits for a person. | R0 |
| L1, prepare | Also draft. | R1 |
| L2, act inside | Also make reversible changes in your own systems. | R2 |
| L3, act outward | Also reach known people and publish. | R3 |
| L4, operate | The same as L3 for now. R4 always waits. | R3 |
Every agent starts at L0. A bare task always runs at L0.
Asking for more, never less
On an agent you can tick tools that should always ask a person first, whatever their class. A rule can only add approval; nothing you set lets an agent skip one its level requires.