At TechXchange on October 27, IBM presented a coordinated approach to AI agent identity, access control, and evaluation. Its availability table lists Vault 2.2, Agent Identity, and watsonx Orchestrate Evaluate capabilities as generally available. Orchestrate Discover and Contain remain in tech preview. That distinction prevents every demonstrated function from being treated as equally ready for deployment.

Accountability follows each step

The separate Agent Identity announcement is dated October 26. It describes a central inventory, agent ownership, delegated authority, and short-lived tokens scoped to a task. Verify supplies identity and policy controls, while Vault complements the architecture with target-specific credentials for systems unable to consume delegated tokens. A structured trail connects the person, agent, and downstream action. These are described product capabilities, not our audit of a working customer environment.

In our example, an employee requests an order summary. The agent may be allowed to read its entries, but that does not imply permission to change the recipient or issue a refund. A useful acceptance scenario would trace whether the request reaches the correct tool and whether an additional action requires separate authorization. This example is not an identified weakness in an IBM product.

What should remain verifiable

Our interpretation is that naming an agent and limiting its actions address different problems. An inventory helps identify the responsible team, but does not itself establish that every external service enforces the same rule. We would inspect the entire path to the final system, including refusals and the case where a user withdraws the task.

Revocation also deserves a separate test. We would distinguish preventing future requests from an operation already completed. If a change has happened, it should remain visible in the record, together with any necessary subsequent human decision. We are not claiming that revoking a token can automatically reverse a business operation.

A bounded pilot before expansion

In practice, we would begin with a low-risk process, clear inputs, and a predefined set of permitted actions. We would include an expired authorization, an incorrect recipient, and interruption between two steps. The useful outcome is evidence showing who requested the operation, its scope, and what actually happened, including failures.

The announcement matters because it attempts to connect agent governance with existing enterprise systems. A generally available component does not turn another component's tech preview into a completed guarantee for the whole architecture. SIExpo has not installed these products, created agents, or independently confirmed promised deployment outcomes.

*AI illustration: a conceptual editorial image, not a documentary photograph or an actual product screen.*

*Editorial verification: 2026-10-29T07:14:17.195221+00:00.*