← Insights
Agents
What Your AI Agent Is Allowed to Do: APIs, Guardrails, and Escalation
Agents should act only through approved APIs, within written limits, and hand off to a person when unsure. Our operating rules for agents, stated in public.
For: IT lead or security lead · 1 min read ·
Start with a permission list
Before discussing models, write down what the agent may do. Read a HubSpot contact? Draft an email in Outlook or Gmail? Create a Salesforce task? Update an order status in the ERP? Each action is a separate decision with a separate owner.
Our operating rules
- Agents act only through approved APIs with scoped credentials, never through a person's full login.
- Read access comes first. Write access is added one action at a time, after review.
- Anything customer-facing or financial is drafted for a person to approve, not sent.
- Every action is logged with its inputs, so it can be audited and reversed.
- When the agent is unsure, or a request is outside its list, it escalates to a named person or queue.
Guardrails that actually work
- Allow-lists of actions and records, rather than instructions asking the model to behave.
- Answers grounded in approved sources such as your handbook or policy library, with references shown.
- Limits on volume and value, such as no refunds over a set amount without approval.
- A test set of real and adversarial requests, re-run whenever prompts or models change.
What we do not build
We do not build black-box autonomous systems that act on their own judgment across your business. An agent is a narrow tool with a written job description, and a person stays accountable for its work.