Straight answers.

If every agent you run lives on Azure and only ever uses Azure models, Foundry may be enough, and we'll say so on the call.

Most enterprises run agents on more than one cloud and framework. A hyperscaler governs its own cloud. AgentOven governs all of them from one control plane, including agents running on Azure.

They are strong runtimes. AgentOven routes to their models and sits above agents hosted on them, so you get one policy, audit and cost layer instead of a separate console and permission model per cloud.

They will build governance for their own cloud. Governing a competitor's cloud works against their business model, so a neutral layer across clouds is the part they are least likely to build.

Yes. AgentOven is self-hosted: run it in your VPC, on-premises or fully air-gapped. Prompts, traces and data stay inside your boundary.

No. You register existing agents as they are, whether they are built with LangGraph, CrewAI, the OpenAI SDK or your own code. Guardrails, cost tracking, approvals and audit apply in the control plane, and the SDKs are optional.

No. The control plane is a single binary backed by PostgreSQL. Run it with Docker Compose, on a VM, or on Kubernetes with the provided Helm charts if you prefer.

It overlaps with LangSmith on tracing and adds governance, open protocols and world-based testing. Traces are exported over OpenTelemetry, so your existing Grafana, Datadog or Jaeger setup keeps working.

The platform or AI-enablement team runs it. Security and compliance teams use the audit trail and approvals. Product teams keep shipping agents in the framework they already use.

Agents and tools connect to AgentOven over open standards, A2A for agent-to-agent and MCP for agent-to-tool, instead of a proprietary SDK. You can change frameworks, models or providers without rewiring every agent.

ACP support is on the roadmap and not shipped yet. When it lands it will run through the same policy, metering and audit path as A2A and MCP.

Worlds are in early access. A world is typed, simulated state, such as orders, customers and accounts, that an agent's tools actually change. A scenario adds simulated users, injected faults and facts that stay hidden until the agent asks.

Most evals ask a judge model whether a response looked right. AgentOven grades the world afterwards, for example whether a refund stayed within the order total, so the same run gets the same score every time.

Not automatically yet. Promoting an agent version to production requires a human approval, and scenario results (early access) give the approver evidence. Automated scenario gates are on the roadmap.

Users sign in with SSO over OIDC or Microsoft Entra ID with role mapping. Machines use scoped API keys and service accounts. Fine-grained RBAC controls what each person and agent can do, and every action is attributed to a principal in an immutable audit trail.

Yes. Guardrails, including PII detection and prompt-injection checks, run on every call, and cost is tracked per agent. A recipe can hold any risky step, such as a large refund, at a human gate until someone approves it.

Yes. The core is Apache 2.0 licensed: agent registry, model routing, A2A and MCP, workflows, tracing and cost tracking. The enterprise edition adds SSO, cross-org federation, worlds and scenarios, extended retention and support.

AgentOven Cloud, a fully managed and usage-based version, is coming soon and is not generally available yet. You can join the waitlist on the pricing page for early access. Today you can self-host the open-source edition or run the enterprise edition on-premises or in your own cloud account.

Enterprise is a flat platform license rather than a per-token charge. Book a demo and we will size it to your agent count and deployment model.