WSO2 has announced the general availability of WSO2 Agent Manager, an open-source platform designed to provide centralized governance, identity management, security controls, and operational oversight for AI agents running across different models, frameworks, and deployment environments.
The release addresses a growing enterprise challenge: while organizations can build agents using an expanding range of models and frameworks, the infrastructure used to manage their identity, permissions, behaviour, and lifecycle is often fragmented. WSO2's approach separates agent governance from agent logic, allowing common controls to be applied independently of the underlying model, framework, or runtime.
Agent Manager entered beta in June 2026. The general availability release adds deeper agent identity capabilities, governance controls for Model Context Protocol (MCP) interactions, and a Kubernetes-native sandboxed runtime. The platform is designed to run across cloud, on-premises, and hybrid environments, to allow organizations to manage agents without tying governance to a single AI provider.
A central focus of the release is agent identity and authorization. Unlike conventional applications, agents can invoke tools, access APIs, delegate work, and interact with other agents, creating additional questions around authority and accountability.
Agent Manager provides a central inventory and management layer, with capabilities including verifiable agent identity, role-based access, delegation, token exchange, and access revocation. It also includes lifecycle controls for moving agents through development, staging, and production, as well as the ability to suspend agents when required.
The platform provides more than 40 built-in controls covering areas such as personally identifiable information masking and rate limiting. These policies can be applied across different parts of an agent workflow, including the agent, MCP, and LLM layers.
The general availability release also introduces a sandboxed execution runtime, reflecting growing concerns about agents being given access to files, tools, APIs, and enterprise systems. The Kubernetes-native runtime is intended to provide a controlled environment for executing agents, while allowing their activity to be monitored and managed.
Agent Manager also uses OpenTelemetry for tracing and includes evaluation capabilities designed to monitor agent behaviour over time. Rule-based and LLM-based evaluations can be used to identify issues such as unexpected token consumption, behavioural changes, or declining response quality.
WSO2's approach is built around framework and model independence. The platform supports technologies including LangChain, CrewAI, Amazon Bedrock, Azure, Ballerina, and custom-built agents, while using standards and technologies such as OpenTelemetry, MCP, and OAuth 2 extensions.
The underlying idea is that governance should remain separate from the AI technologies being governed. Models, frameworks, and providers can change rapidly, and organizations may use different combinations for different workloads. A separate governance layer could allow teams to change those underlying technologies without rebuilding identity, policy, and monitoring controls each time.
The wider industry conversation suggests that agent governance is increasingly becoming a platform problem rather than simply an AI development problem. AWS, Microsoft, and other providers are all adding combinations of non-human identity, tool authorization, policy enforcement, runtime isolation, observability, and evaluation to their agent platforms. Recent industry analysis from Qovery has also argued that no single platform yet provides a complete governance stack across every layer, with enterprises likely to combine identity systems, policy engines, infrastructure control planes, and isolated runtimes. That makes the idea of a framework-independent control plane increasingly relevant, particularly for organisations that expect their agent estate to span multiple models and cloud platforms.
The emerging challenge is therefore similar to earlier shifts in cloud and platform engineering: workloads may use different technologies, but organizations still require common controls around identity, security, policy, observability, and lifecycle management. The question is no longer simply how enterprises build AI agents, but how they consistently identify, govern, and control them once they are operating at scale.