WSO2 Agent Manager Is GA: Open-Source Control Plane for AI Agents

Your company has somewhere between 20 and 200 AI agents in production. You’re not sure of the exact number — the data team spun up three last month, someone connected a CrewAI workflow to your CRM in August, and IT isn’t certain if the evaluation agents from Q2 are still running. Gartner calls this agent sprawl. And it’s about to get much worse: the firm projects the average Fortune 500 enterprise will be managing over 150,000 agents by 2028, yet only 13% of organizations believe they have adequate governance in place today.

WSO2 shipped Agent Manager as generally available this week — an Apache 2.0 open-source control plane designed to govern AI agents across frameworks, models, and deployment environments. The idea isn’t to replace your agent frameworks. It’s to sit above them and enforce the identity, policies, and observability that your security and compliance teams keep asking for, without requiring you to rebuild your agents from scratch.

Not Another Framework. A Control Plane.

The distinction matters. WSO2 Agent Manager is not another agent framework competing with LangGraph or CrewAI. It’s the governance layer that sits above your existing frameworks and makes them enterprise-ready. It works with LangChain, CrewAI, Amazon Bedrock Strands, and Microsoft Agent Framework out of the box — and because it’s built on OpenTelemetry, OpenAPI, and MCP, custom implementations hook in without rewriting existing code.

The Apache 2.0 license means you self-host with full data sovereignty, or use the managed SaaS offering with a 99.99% SLA. That’s a meaningful choice in a space where most governance tooling locks you into a specific vendor’s runtime. LangSmith is excellent — if your entire stack is LangGraph. Temporal handles durable execution well — but it’s not a governance layer. Agent Manager is trying to be the framework-agnostic control plane the ecosystem has been missing.

Agent Identity as Infrastructure

The most important architectural decision is the identity model. Instead of agents running under shared human service accounts — a security nightmare as agent counts scale — Agent Manager gives every agent a verifiable machine identity using an OAuth 2 extension for MCP. Role-based access, token exchange, and delegation are built in. Every tool call, every agent-to-agent interaction, every external API call is authenticated, authorized, and auditable. When something goes wrong at 2 AM, you know exactly which agent did what, under what permissions, and you can revoke access in real time.

40+ Guardrails, Four Levels Deep

The policy enforcement system operates at four levels: organizational-wide rules, per-agent policies, MCP tool-call controls, and LLM input/output filtering. The 40+ built-in guardrails are mapped to the OWASP Top 10 for LLMs, covering PII masking, URL validation, semantic prompt validation, rate limiting, and cost controls. If an agent goes rogue, there’s a single-click suspension capability. Importantly, these guardrails run at the infrastructure layer — your developers configure policies; they don’t write guardrail logic inside every agent.

That said, 40+ configurable guardrails also means 40+ guardrails to configure. The real test for teams adopting this will be whether the defaults are sensible enough to deploy without a week of policy tuning, and whether the enforcement is granular enough to avoid blocking legitimate agent behavior. WSO2’s Chief AI Officer Dr. Rania Khalaf frames this as a solved tension — “speed and control shouldn’t be a tradeoff” — but every team will stress-test that claim differently.

Zero-Code Instrumentation

The observability story is handled by the amp-instrumentation Python package and a Kubernetes init container that auto-instruments existing agents without code changes. You get full OpenTelemetry traces, metrics, and logs across your entire agent fleet, plus 24 built-in evaluations for ongoing production monitoring. For teams already on Kubernetes, the quick start is a single Docker command:

docker run --rm -it --name amp-quick-start   -v /var/run/docker.sock:/var/run/docker.sock   --network=host   ghcr.io/wso2/amp-quick-start:v1.0.0

Zero-code instrumentation means you don’t need to refactor production agents to get observability. That’s the right call for enterprises with agents already in the field that weren’t built with governance in mind from day one.

The Missing Layer

The agent governance gap is real — Gartner’s numbers suggest 40% of agentic AI projects will be cancelled by 2027 due to cost, unclear value, and insufficient risk controls. A control plane that decouples governance from agent logic — Apache 2.0, framework-agnostic, self-hostable — is a credible answer to that problem. The GitHub repository is the starting point; the managed SaaS trial is at console.agent-manager.cloud.wso2.com. If you’re running multiple agent frameworks across teams and can’t answer basic questions about what your agents are doing in production, Agent Manager is worth a serious look.