WSO2 Agent Manager GA: an agent control plane you host
WSO2 Agent Manager hit GA September 15 under Apache 2.0, self-hosted or SaaS. What an open agent control plane changes about governing the agents you already run.
Every shop we talk to has more agents than it has records of agents. One in a Slack workflow, one in a support queue, two in a repo somebody forked in June. WSO2 Agent Manager went generally available on September 15 as an open agent control plane — Apache 2.0, self-hosted or managed SaaS — and the interesting part is not the feature list. It is that the governance layer for your agents is now something you can run on your own cluster instead of rent from the vendor whose agents you are trying to govern.
What actually happened
WSO2 announced general availability after a beta that opened May 5. The product sits above your agents rather than inside them, and it is deliberately framework-agnostic: the release names LangChain, CrewAI, Amazon Bedrock Agents and Microsoft Agent Framework, across Python and Ballerina runtimes.
What ships in the box, per the GA release: 40+ built-in guardrails including PII masking and rate limiting, verifiable agent identity with role-based access, governance at the MCP layer rather than only the model layer, a sandboxed Kubernetes-native runtime, OpenTelemetry-based observability, and continuous evaluation monitoring rather than a one-time eval run.
The source repo confirms the mechanics: Apache 2.0, Helm charts for the control plane plus build and observability extensions, and OpenTelemetry instrumentation for AI frameworks that the project describes as requiring zero code changes. It monitors agents it hosts and agents deployed elsewhere. WSO2 frames the self-hosted option as sovereignty over where agent data lives and runs — which is the correct framing, because that is the part a SaaS console cannot give you.
Why an open agent control plane matters for your business
Most teams govern agents at the wrong layer. They write a policy about which model is approved, then hand every agent an MCP server with write access and no identity of its own. The model was never the risk surface. The tool calls are.
A control plane you host changes three practical things. Agents get identities, so a log line says which agent did it instead of which API key did it. MCP-level policy means you can allow a read server and deny a write one without editing every agent's config. And OpenTelemetry means agent traces land in the observability stack you already pay for, not a second dashboard nobody opens.
The honest caveat: this is infrastructure, and infrastructure has a cost. Running Agent Manager means running Kubernetes, Helm releases, and an upgrade cadence. If you operate three agents, a spreadsheet and scoped credentials are cheaper and just as effective. If you operate thirty, you already have sprawl and you are just not measuring it yet.
The strategic read is the license. Apache 2.0 with a self-hosted deployment path means your governance layer does not become a second lock-in on top of your model lock-in. That is worth more than any single guardrail in the list.
Key takeaways
- WSO2 Agent Manager reached GA September 15, 2026 under Apache 2.0, deployable self-hosted or as managed SaaS
- It governs agents across frameworks — LangChain, CrewAI, Bedrock Agents, Microsoft Agent Framework — rather than locking you to one
- Policy lands at the MCP and agent-identity layer, not just the model layer, which is where the real risk sits
- OpenTelemetry instrumentation means agent traces go to your existing observability stack
- Count your running agents before you buy a control plane; under a handful, scoped credentials and a list are cheaper
Do you know how many agents your business is currently running? We inventory what you have, give each agent its own identity and scoped tool access, and wire the traces into logs you can actually read. See how we build AI agent systems, or send us your setup and we'll map it.
Sources: WSO2 GA announcement, wso2/agent-manager on GitHub, WSO2 beta announcement.
- #ai-agents
- #governance
- #open-source
- #mcp
- #observability
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenAI's misalignment disclosure now runs on a 6-day clock
OpenAI published its AI misalignment disclosure framework with six new incident reports and stated review windows of 6 and 12 business days. Put the number in your contract.
Read itHang Ten's $85M bets AI consulting is a skills library
Vishal Sikka's Hang Ten Systems raised $53M more for enterprise AI consulting built on packaged, reusable AI skills. The model works at any size — including yours.
Read it