IBM adds identity, kill switches and cost tracking to rein in AI agents
At TechXchange, IBM introduced tools to discover AI agents across platforms, give each one a verified identity and least-privilege credentials, sandbox or kill a misbehaving one, and track what each agent costs.
IBM used its TechXchange conference on October 27 to put governance tools around AI agents, the software that acts on its own across a company's systems. The release covers four jobs most companies have not solved: finding every agent that exists, giving each one a verified identity, holding it to least-privilege access, and testing what it actually does. "The winners in enterprise AI are changing how their business runs, not simply adding more AI to it," said Dinesh Nirmal, Senior Vice President of IBM Software.
What the tools do
IBM extended HashiCorp Vault, which it owns, to issue agents short-lived credentials scoped to each agent's needs, and to revoke access in real time when an agent tries something it should not. A companion piece, Agent Identity, keeps a record of each agent's owner, lifecycle status and granted authority, so an agent is a named thing with an accountable owner rather than an anonymous script with a shared key. Vault 2.2, now generally available, gives agents their own secured accounts with auto-renewed credentials.
IBM's watsonx Orchestrate adds three controls across platforms. Discover, in tech preview, finds agents built in third-party environments including Amazon Bedrock, Microsoft Foundry and Google Gemini. Contain, also in tech preview, runs an agent in an isolated sandbox and adds a kill switch. Evaluate, which is generally available, tests agents, models and MCP servers for quality, accuracy and compliance, and tracks their costs. IBM did not publish pricing.
Why it matters
The gap these tools name is the one most companies will hit next: agents created faster than anyone is tracking, each holding credentials nobody scoped, running work nobody priced. That is where AI spend quietly turns into both risk and waste, because an agent you cannot see is one you cannot cost or switch off. Whether or not a company buys IBM's version, the capabilities are the checklist worth holding any agent deployment to: know which agents exist, scope what each can reach, be able to stop one, and measure what each costs against what it produces. An agent that can be named, limited, halted and priced is one you can judge on value per dollar. One that can't is a line item with no owner.