IBM previews tools to discover and isolate AI agents across platforms
IBM has put watsonx Orchestrate’s agent discovery and containment capabilities into tech preview. The tools connect to other AI platforms to find agents and provide isolated execution with a shutdown control. The preview extends management across platforms, alongside identity and credential products whose availability is separate from the new discovery and containment tools.
Artificial Intelligence··Morning
Discovery reaches agents built elsewhere
IBM announced technical previews for discovery and containment in watsonx Orchestrate, its AI-agent management system, at TechXchange on 27 October. The previews extend oversight to agents operating across different platforms. These software agents carry out tasks using models and connected tools; they can therefore require access to business applications and data.[1], [2]
The discovery tool connects to Amazon Bedrock, Microsoft Foundry and Google Gemini to find third-party agents instead of relying on manual registration. Its scope also covers models and Model Context Protocol servers, which provide connections between models and tools or information. IBM lists discovery as a tech preview in its availability schedule.[1]
Containment isolates execution and adds a stop control
The containment preview runs agents in isolated sandboxes and gives teams a control for shutting an agent down. IBM presents that separation as protection against one agent interfering with others. Containment remains in tech preview as well. The company describes a central management experience for agents built elsewhere, alongside its existing tools for testing quality, accuracy and compliance and tracking model and token use.[1]
Identity and credential tools have separate availability
HashiCorp Vault, IBM’s system for managing application credentials and secrets, supports expiring access and revocation. Agent Identity supplies ownership, lifecycle status and granted authority. IBM says their combined controls can deny access and revoke a session when an agent attempts an action outside approved permissions. Vault 2.2, Agent Identity and evaluation capabilities are generally available. Their status is separate from the discovery and containment previews; the announcement supplies no independent enforcement test.[1]
