WSO2 has released Agent Manager as a generally available open-source control layer for organizations running AI agents across different frameworks and infrastructure. The platform separates governance from agent code so teams can change models or runtimes without rebuilding identity, monitoring and policy controls for each one.
The release adds verifiable identities for agents, role-based access, delegation, token exchange and revocation. Teams can track agents through development, staging and production or suspend them when necessary. More than 40 built-in controls cover areas such as masking personally identifiable information and limiting request rates, with policies applicable at the agent, language-model and Model Context Protocol layers.
A Kubernetes-native sandbox provides an isolated place for agents that can touch files, tools and enterprise APIs. OpenTelemetry traces activity, while rule-based and model-based evaluations watch for changing behavior, unexpected token use and declining response quality. Supported technologies include LangChain, CrewAI, Amazon Bedrock, Azure, Ballerina and custom agents. General availability follows a beta launched in June. No single layer eliminates agent risk, and organizations may still need separate identity systems, policy engines and infrastructure protections. WSO2’s release addresses the coordination problem by giving those controls a common inventory and lifecycle surface rather than tying them to one AI provider.