IBM strengthens governance of AI agents in enterprises
IBM enhances watsonx Orchestrate with an AI gateway, a governed catalog, and agent identity control. The goal: to give companies a central view of an increasing number of agents before the stack becomes unmanageable.
IBM announced at its TechXchange conference a series of tools designed to oversee artificial intelligence agents in enterprises. The centerpiece is a new AI Gateway integrated into watsonx Orchestrate, IBM's control layer for managing AI agents, third-party models, and MCP servers. In a context where deployments are multiplying, the challenge is no longer just to make these agents work, but to know who acts, with what rights, and on which data.
Quand les agents se multiplient, la gouvernance devient un métier à part entière
IBM addresses here a very concrete problem: the proliferation of AI agents in enterprise systems. The AI Gateway is intended to discover and manage third-party agents, models, and Model Context Protocol (MCP) servers, the protocol used to connect tools and data sources to an agent. In other words, IBM is trying to restore some order in a mechanism that can quickly resemble an overheating engine room.
The watsonx Orchestrate catalog goes further by allowing the discovery, sharing, evaluation, and reuse of approved agents and tools. For certain MCP servers provided by Box, Visier, and New Relic, IBM also displays precise security signals: protocol compliance, tool schema integrity, description security, input validation, and rejection of default credentials. This type of control matters because a poorly designed tool description can become an open door, even without bad intent.
Tester, tracer, identifier: l’autre moitié du problème
IBM has also added testing and observability functions. Teams can now customize the criteria used by a large language model to evaluate a process, define and simulate test paths, monitor workflow execution with Trace Inspectors, and use an agent called AgentOps Agent, responsible for examining other agents, generating test cases, and automatically optimizing performance.
At the same time, IBM has made IBM Agent Identity generally available, a control plane that gathers registration, delegation, rule enforcement, and governance of AI agents. This component allows maintaining a central inventory, documenting ownership and metadata, classifying risk, archiving authorized tool connections, and keeping audit trails. For CIOs, this means applying to agentics what companies have ended up imposing on human identities: knowing who can do what, and within what limits.
Le vrai sujet: éviter qu’un agent “réussisse” deux fois
The issue is not theoretical. Recent work on agent reliability shows that a system can believe it has completed an action while it has caused a duplicate effect. If a support agent retries a refund because it did not receive confirmation, the second submission can create a new refund. The risk is not abstract: it lies in the fact that a non-returned result proves neither failure nor success.
In this context, orchestration and governance mechanisms become an operational safeguard, not just a dashboard. Companies must distinguish the authorization to execute an action from the authorization to replay it in case of uncertainty. IBM is therefore pushing its offering towards a complete control layer, with the idea that at large scale, the real difficulty is not launching agents, but properly containing them.
Ce que cela change pour les entreprises
IBM's move reflects a broader market trend: companies no longer want just "intelligent" agents; they want manageable agents. Dinesh Nirmal, senior vice president for IBM Software, told TechXchange attendees that it is crucial for IT teams not only to observe agents for their performance, quality, and cost but also to manage them via a control plane. The message is clear: without governance, the bill always ends up coming due.
For organizations already equipping themselves with agentics, the question is no longer whether a control framework will be necessary, but which one. IBM is betting on watsonx Orchestrate to play this role in distributed environments, with third-party and in-house agents. It remains to be seen if companies will keep pace as promised, since the real constraint will often be less software than organizational: skills, responsibilities, procedures, and the ability to decide quickly when an agent blocks.
By the numbers
- 3 families of capabilities — governance, testing, and observability, announced at TechXchange.
- 3 vendors mentioned — Box, Visier, and New Relic for certain MCP servers.
- 1 control plane — IBM Agent Identity, now generally available.
- 1 conference — IBM TechXchange, where the announcements were made.