Govern the Agents You Did Not Build

· 6 min read agentsgovernanceidentity

Navy and amber hero graphic with bold text: Govern the Agents You Did Not Build

IBM made the most honest admission in agent governance today, and it came inside a product announcement. At TechXchange 2026 in Atlanta, IBM announced that watsonx Orchestrate will discover, contain, and evaluate agents built on other vendors’ platforms, and that HashiCorp Vault will issue those agents governed identities and short-lived credentials. In IBM’s own write-up, the pitch is explicit: manage agents “regardless of where those agents were built or run.”

That is the right problem. It is not the problem most control planes were built to solve. A control plane that only governs agents built on its own platform governs the demo estate, not the real one.

The facts first, with availability stated plainly because it matters. Discover is in tech preview: Orchestrate connects to Amazon Bedrock, Microsoft Foundry, and Google Gemini to find agents built there, and the discovery extends to models and Model Context Protocol servers. Contain is also in tech preview: agents run in isolated sandboxes, with a kill switch to shut one down. Evaluate is generally available: test quality, accuracy, and compliance before deployment, monitor in production, and track cost by agent and token use. On the identity side, Agent Identity combined with Vault is generally available. Agent Identity holds the authoritative record for each agent: who owns it, its lifecycle status, and the authority it has been granted. Vault 2.2 gives agents their own secured accounts, renews credentials automatically, and issues short-lived credentials scoped to the job. If an agent reaches outside its approved permissions, access can be denied and the session revoked in real time.

IBM grounds the urgency in a Gartner figure it cites in the release: by 2028 the average Fortune 500 enterprise will use over 150,000 AI agents, and only 13 percent of organizations believe they have the right agent governance in place. Treat the 150,000 as a vendor-cited forecast, not a count. The 13 percent matches what I see. The governance is thinner than the estate.

I have argued the pieces of this before. Every Agent Needs an Identity made the case that an agent with no named identity cannot be owned, scoped, or shut down. Standing Privilege Is Dead argued for credentials that are issued just in time and expire. The Passport Is Not the Decision warned that a signed credential proves a scope was authorized, not that the scope was narrow or current. IBM’s announcement puts those three pieces into one stack and points the stack at foreign agents. That combination is the news. The tests below are how I would judge it, and they apply to any vendor making the same claim, including a control plane I build myself.

First test: coverage without registration. Discovery that finds an agent only after someone registers it is an inventory form, not discovery. Ask what the connector actually enumerates on Bedrock, on Foundry, and on Gemini, and what it misses: agents running under a developer’s personal account, agents embedded inside a SaaS product, local agents on workstations, and MCP servers stood up by a team for one project. IBM says discovery extends to models and MCP servers, which is the right scope, because the server an agent calls is part of its real authority. Get the miss list in writing. A partial map with strong controls on the mapped part is still a partial system.

Second test: enforcement that travels. A registry entry does not stop anything. What stops an agent is a point it must pass through: a credential it must fetch, a gateway it must call through, a runtime it cannot leave. Vault is the strongest part of this announcement for exactly that reason. An agent built anywhere still has to get a credential from somewhere. If it gets a short-lived, scoped credential from Vault, the limit travels with the work, and revocation lands in real time instead of at the next audit. The kill switch and the sandbox matter the same way, if they act on the foreign runtime and not just on agents Orchestrate hosts itself. In the demo, ask for a denial, not a dashboard. Have the agent attempt an action outside its scope on a platform IBM does not own, and watch whether the action fails, the session ends, and the record shows both.

Third test: read the availability labels as part of the architecture. The cross-platform discovery and containment in this announcement are tech preview. Identity, credential control, and evaluation are generally available. That split tells you what you can build policy on today and what you can only pilot. Write the policy against the generally available controls. Pilot the preview controls with a named population and a review date. Do not let a roadmap item carry a compliance claim. I made a version of this point in The Blocker Is Not the Control Plane: one component, however good, is not the system of inventory, ownership, controls, and audit that governance requires.

There is also a question no vendor announcement will answer for you. If one vendor’s control plane discovers and grades every other vendor’s agents, who grades the control plane? The catalog in IBM’s write-up surfaces evaluation data and risk signals at the moment of discovery. Useful, and worth checking like any other record: what criteria, whose test set, and how old is the score? Cross-platform governance concentrates a new kind of authority in the plane itself. That authority needs an owner, a scope, and an audit trail, same as any agent.

None of this requires waiting for IBM, or for any single vendor to finish the job. The sequence works with the tools already in most shops. Issue identity before access, from one authoritative record, whoever runs it. Make credentials short-lived and scoped at issuance, because issuance is where the real decision sits. Point discovery at the platforms your teams actually use, and keep a written list of what discovery cannot see. Then test enforcement where the agent runs, not where the console lives.

Your teams will keep building agents on whatever platform fits the job in front of them. That is not a failure of discipline. It is how enterprise technology has always spread. Governance that demands a rebuild before it will protect the work will be routed around, and it will deserve to be. Govern the agents you have, including the ones you did not build and would not have picked. That is the whole job now.