Permissions
Access should match responsibility. We design around role-based access, company versus user scope, and least privilege.
- Define what a user or agent can see
- Define what tools it can call
- Keep sensitive actions scoped
Human-controlled automation
Useful automation should still be accountable.
Our security approach starts with the workflow: scoped access, explicit approval points, and a visible record of what happened.
Human review
Not every action should be automatic. Draft, review, approve, and execute can remain distinct steps.
- Draft first where appropriate
- Route exceptions to people
- Require approval before sending
Integration security
APIs and MCP tools should be connected intentionally, with credentials and capabilities limited to the workflow.
- Scoped tools and credentials
- Clear integration boundaries
- Test with realistic permissions
Auditability
Teams should be able to understand what started a workflow, what data was accessed, what the agent did, and who approved it.
- Initiator and agent activity
- Data and action visibility
- Approval history where supported
Customer data
Data handling should be discussed before implementation, including storage, model access, retention, and training policy.
- Document the data path
- Confirm provider behavior
- Match retention to the use case
Honest scope
We will not imply certifications, guarantees, or controls that have not been verified for the specific implementation.
- Review the actual stack
- Record open decisions
- Update controls as systems change
Security is part of the design
Bring the sensitive workflow.
We can map the data, tools, roles, and approval points before deciding whether automation belongs there.