Skip to main content

Authority should be designed, not assumed

An action layer is useful because it can affect the product. That is also why its authority must be explicit, narrow, and observable. Dubot’s intended governance model begins with named capabilities and agreed control points. It does not begin with unrestricted product access.

Initial rollout safeguards

Bounded scope

The initial rollout covers a small number of reviewed customer outcomes.

Explicit confirmation

Sensitive or consequential steps can require customer or agent approval, and actions that write always confirm.

Known completion

Each journey defines evidence of success, failure, or safe stopping.

Visible recovery

Unexpected state should lead to a clear stop or handoff, not hidden improvisation.

How actions are governed

Every action carries a policy that decides how it may run. Actions that write to the product always confirm. Read-only actions can be allowed to run automatically when a team is comfortable with their scope.

What the action layer can see

Through a connected product, Dubot sees the shape of your product’s own API responses, the same information a developer would see with the browser network tab open while using the product. It does not receive hidden or privileged access beyond what the connected surface already returns.

How support data is handled

When a support conversation is used to measure or improve deflection, that data is read-only and kept to a minimum footprint. Conversation content should be de-identified or reduced to outcome-level signal, according to what the customer’s security team prefers. Connector scopes should be requested at the minimum needed for a specific use case, and narrowed when a smaller scope is enough.
Actions are reviewable before use, in the Actions interface or over MCP, so a security or engineering team can inspect exactly what an action does before it is attached to an agent.

Technical review checklist

Before customer use moves beyond a controlled test, Dubot and the customer should document:
  • environments and domains in scope;
  • user identity and session assumptions;
  • credentials or tokens, if any, and how they are stored;
  • the minimum data passed into and out of a journey;
  • actions that are allowed, confirmation-gated, or prohibited;
  • logging, retention, and access expectations;
  • failure, recovery, and support ownership;
  • testing and release approval.

Claims that require verification

This documentation does not make claims about certifications, encryption standards, data residency, subprocessors, penetration testing, uptime, or contractual controls. Those details will be published only when they have been verified against current product and company evidence.

A practical evaluation principle

The first question is not “Can the system click anything?” It is “Can this one approved customer outcome be completed safely, predictably, and visibly?”