Companies govern the people who work for them through four things: authority granted, limits set, oversight exercised, and records kept. Nobody hands a new hire the company card on their first morning with no limit and no receipts.
AI agents joined the workforce with none of that structure. They issue refunds, approve discounts, pay vendors and push code, and the question most tools can answer about them is whether they may call a tool at all.
That is not the question a business asks. A business asks whether this agent may refund this customer this amount, given everything that has already happened today. The honest answer is rarely yes or no. It is: authorized, up to 150.00, and above 100.00 a person has to say yes.
Shepherd is the structure that gives that answer. Your agents ask before they act, they are told what they are authorized to do and where the limits fall, and every answer is signed into a record you can show somebody.
Four things an organization does
- Authorize
- Rules, limits and approvals, written once and applied to every agent on every surface you run.
- Oversee
- A console where a person sees what needs deciding, what the agents did, and what they are authorized to do.
- Prove
- A signed, hash-chained record of every decision and every outcome, which an auditor can check without taking our word for it.
- Adapt
- Replay a rule change against what really happened, and watch it run beside the live rules, before it governs anything.
What we hold ourselves to
- 1
No model decides.
A model can help a person write a rule. Only a rule a named person accepted is enforced, and the same rule with the same history gives the same answer every time.
- 2
What an agent was told can only tighten.
Something an agent heard in a conversation can lower a limit. It can never raise one.
- 3
When in doubt, the answer is no.
If Shepherd cannot be reached, a consequential action does not go ahead unless you chose otherwise in advance, and the record says which happened.
- 4
Limits, not rewrites.
Shepherd tells an agent what it is authorized to do, and the agent applies it. Shepherd does not quietly change what an agent asked to do.
- 5
The record is of what happened.
Budgets count what agents report they did, not what they meant to do. Where the two cannot be told apart, the larger figure is counted, because the smaller one would quietly widen the next answer.
- 6
We are not your database.
What a decision needs to know about a customer comes from your own systems at the moment of deciding. The record keeps what was decided and why, never secrets or raw personal data.
What we do not promise
Shepherd does not stop every mistake. It bounds what an agent is authorized to do and records what it did, which is a different and smaller promise. We would rather make that promise and keep it.
Start by watching. Connect an agent, and Shepherd records what it does before a single rule applies. Grant authority when you have seen enough to know what to grant.
Download for