At TechXchange 2026 in Atlanta, IBM announced agent identity, credential and evaluation controls built around HashiCorp Vault and watsonx Orchestrate, aimed at letting enterprises trace and limit what AI agents can do.
IBM used its TechXchange 2026 conference in Atlanta to announce a set of controls for AI agents working across business systems, built around verified identities, least-privilege credentials and a single place to discover, secure and evaluate agents wherever they were built.
The centrepiece is HashiCorp Vault Agent Identity, which creates an authoritative record for each agent covering who owns it, its lifecycle status and the authority it has been granted. Vault issues short-lived credentials limited to what each agent needs, and access can be denied and the session revoked in real time if an agent tries to act outside its approved permissions.
IBM watsonx Orchestrate adds discovery across Amazon Bedrock, Microsoft Foundry and Google Gemini, sandbox containment with a kill switch for a rogue agent, and evaluation of quality, accuracy and compliance before deployment and in production, with cost tracking by agent, model and token use. Vault 2.2, Agent Identity, Orchestrate evaluation, IBM Concert and HashiCorp Terraform policy are generally available, while Secure Hub and several other pieces are in public preview or tech preview, and the agent tools integrate with NVIDIA OpenShell secure runtime software.
The announcement itself is IBM's own release, carried by AAP, so the capabilities and availability claims rest on the company's word. The eye-catching scale figures in the release, that the average Fortune 500 enterprise will use more than 150,000 AI agents by 2028 while only 13 percent of organisations believe they have the right governance in place, are IBM's citation of Gartner and were not independently re-checked here, and the release names client speakers rather than giving measured deployment results, so this story claims no proven customer outcome.
IBM software chief Dinesh Nirmal framed the point bluntly, saying the winners in enterprise AI are changing how their business runs rather than simply adding more AI to it, and that building trust and control into the operating model is what turns agents from a technology into a way of doing business. The test now is whether the tech-preview pieces, discovery and containment, reach general availability and whether enterprises can audit agent actions as routinely as they audit human access, given IBM itself notes that statements about future direction are subject to change.