IBM announced on October 27, 2026 that Agent Identity is generally available, giving businesses deploying AI agents a new option for managing their access to company systems and tracing whose authority they use.
AI agents are software that can use tools to take actions, not just answer questions. IBM’s offering addresses three practical questions: who owns the agent, what may it do on someone’s behalf, and what record remains of its actions?
The release does not make every related safeguard generally available. Separate watsonx Orchestrate features for discovering agents across platforms, running them in isolated sandboxes and shutting them down with a kill switch remain technology previews.
Whose permission is the agent using?
IBM says its Verify software handles agent identity and authorization, including permission to act on someone’s behalf. HashiCorp Vault supplies temporary credentials for particular target systems when those systems cannot directly accept the delegated permission. These are complementary controls: identifying an agent and giving it access are different jobs.
The company also describes audit records that connect identities, permissions, credentials and activity. Those records are intended to help a business reconstruct who authorized an action and how it was carried out; the product description does not establish complete coverage of every integration.
HashiCorp’s documented delegated-access workflow shows how authority can be narrowed. Access within Vault must fit three limits at once: the person’s existing permissions, the maximum permissions assigned to the agent, and the restrictions for that particular session.
The practical implication is that acting for an employee need not mean inheriting everything that employee can do. A broadly authorized person can delegate a narrower task. That conclusion applies to the documented Vault controls, not automatically to every application an agent reaches.
Short-lived access depends on the setup
Issuing temporary credentials is different from storing an existing password. Vault’s lease documentation says dynamically generated secrets have leases that can expire or be revoked, while its ordinary key-value store does not issue leases. Moving a password into that store therefore does not, by itself, make the password expire when an agent finishes a task.
The documented Vault setup calls for a Vault Enterprise or HCP Vault Dedicated server, a supported identity provider and administrator permissions. Those requirements describe the Vault workflow, not the full licensing terms for IBM’s combined offering.
As of October 27, IBM’s Agent Identity product page offers a demo but lists no numeric price. Businesses should confirm costs, included components and support for their intended applications before treating this as an upgrade they already own.
For related context on controlling what agents can do with business tools, see RohitAI’s coverage of Broadcom’s AgentMinder.
A useful first evaluation is one clearly bounded workflow: identify its owner, define the access the task actually needs, establish when credentials expire and check which actions appear in the audit record. That turns the launch into a concrete purchasing question: can the system account for this agent’s work in the applications the business uses?