Control rooms, plant DCSs and pipeline SCADA already automate what must happen in milliseconds. The opportunity for AI agents lies elsewhere: in the hours and days between a signal and a decision, where people are compiling evidence, chasing reports and reconciling numbers.
Set autonomy by consequence
How we set autonomy across an energy agent squad
From our energy agent designs
| Kind of work | Typical autonomy | Who decides |
|---|---|---|
| Grid, plant and pipeline signals | Observe or suggest | Control room, plant head, integrity engineer |
| Exchange bids | Act with approval | A trader submits every bid |
| Billing corrections, regulatory packs | Act with approval | Officer or regulatory head |
| Payment and settlement matching within tolerance | Act within limits | Agent, with exceptions to an officer |
Note: Designs, not results.
Across the 28 agent roles we have designed for energy, none can switch, dispatch, control a plant, submit a bid or approve a dig. Two — a payment reconciler and an exchange settlement reconciler, both matching within tolerance — act on their own; everything else either suggests or prepares an action for a person to approve.
What good governance looks like
- An inventory of every agent, its data and its autonomy level
- Policy limits and a kill switch that halts all agents
- Every plan, tool call and decision logged
- Overrides reviewed and fed back into evaluation
Where to start
Start with agents that observe and draft in one domain, publish their autonomy levels and owners, and widen their scope only as the override rate falls.