AI agent / MCP quickstart

​

1. Keep the actors distinct

AtlaSent distinguishes the organization, initiating human where applicable, agent/workload identity, and downstream executor. A human login does not silently become the agent’s identity or authority.

​

2. Technical access is not organizational authority

An MCP server or tool catalog can make a tool technically callable. That does not establish authority to perform a consequential action.

​

3. Evaluate before the tool effect

For a consequential tool call, submit the protected action and current context to the canonical AtlaSent evaluation path before invoking the effectful operation. Use a sandbox/non-consequential target for the first integration.

​

4. Apply bounded delegation only where it exists

If the organization has explicitly delegated authority to the agent/workload, include the applicable authority context. Never infer delegation from the user’s session, tool availability, or prior successful action.

​

5. Verify the execution artifact

Where the protected action requires a Permit, verify it at the gate immediately before the effect. Material changes to the authorized request require fresh authorization.

​

6. Preserve attribution and evidence

The evidence chain should make it possible to distinguish:

• initiating human/organization provenance
• agent/workload identity
• exact protected action
• authority/delegation inputs
• authorization decision
• Permit issuance/verification
• execution/effect evidence, when actually established

​

Negative case

Give the agent technical access to a safe test tool while withholding the required AtlaSent organizational authority. The action should fail closed at the authorization boundary rather than relying on the tool’s own availability as permission.