See It Work — The 40 agents nobody owned | Book 10 · Scaling AI Agents

Agents are sprawling across finance, ops, and support — half of them built on someone's personal account. Nobody can answer the board's question: who owns what, and what can it touch? Watch the first one get a governed identity.

This is the pilot trap in one image: pilots multiply faster than governance, and the org loses track. The fix isn't a spreadsheet of agents — it's giving each one an identity that self-documents every action.

What this means for you

You can't govern what you can't name. Governance at scale doesn't start with a policy — it starts with identity: bind each agent to a principal, its allowed actions, and an owner, and every action it takes documents itself. The inventory stops being a chase and becomes the exhaust. What this means for you: you can finally answer 'who owns this agent and what can it touch?' — because every agent carries an identity that documents its own actions.

That binding was an agent action — so it left a record. Here's the receipt that turns a shadow agent into a governed one:

Role-Binding Receipt

bound_rolefinance_reviewer

principal_idyour node identity

allowed_actionproduce_forecast_artifact

ownerassigned

attestedtrue

Every governed agent produces this same shape, every time — which is what makes a 40-agent fleet auditable instead of a liability.

For the technical reader — the command, and how to verify it yourself

# one line · you do not need to run this
python examples/load_role_demo.py

jq '.compliance_attestation.summary.action_class' memory/LoadRoleDemo_summary.json
# -> the role's allowed action class

Full step-by-step is in Appendix RX: Hands-On Demonstrations in the book.