AgentWatch
Book a demoEnter console

AgentWatch and Microsoft Agent 365

As of 27 October 2026. Everything said about Agent 365 below comes from Microsoft's product page, microsoft.com/en-us/microsoft-agent-365, read that day. If it has changed, the page wins. Tell us and we will correct this one.

What Microsoft says Agent 365 is

  • Price: $15.00 per user per month, paid yearly (annual commitment). It is also included in Microsoft 365 E7.
  • Licensing: Microsoft recommends a licence "for all users who interact with, own, manage, or sponsor Agent 365–managed agents."
  • Which agents: the registry covers Microsoft's agents, "agents from our ecosystem partners, and any agents you register yourself". Agents "built outside these environments will require additional steps" to appear in the inventory.
  • Where it runs: the page doesn't mention an on-premises or air-gapped option. It says "Existing Microsoft cloud subscribers can sign in to get started."

Where Agent 365 is the better choice

  • Your agents are Microsoft agents. If they are built in Copilot Studio or Microsoft 365 Copilot and your identity lives in Entra, Agent 365 sits in the stack you already run. AgentWatch does not read agent identities from Entra. Its Azure Foundry connector can sign in with Entra ID, and that is as far as it goes.
  • You are buying Microsoft 365 E7 anyway. Agent 365 is included, so the marginal cost is zero.
  • Distribution. Your Microsoft account team, procurement path and support contract already exist. Ours don't, yet.

Where AgentWatch is different

Each line links to where you can check it.

  • Any agent, any framework. AgentWatch sees agents through its SDK, through Gate (a proxy for LLM and MCP calls), through Cursor hooks and through connectors (OpenAI Assistants, Azure Foundry, Vertex Agent Engine, Datadog, webhooks). What each source can and cannot see: data sources.
  • Runs where your data is. Cloud, your own Kubernetes cluster, or a cluster with no outbound internet: self-hosted guide.
  • Published coverage. Each OWASP Agentic (ASI) and LLM Top 10 item is marked covered, partial or missing, with a reason per gap, generated from the code: coverage.
  • Published misses. Thirteen real incidents replayed against the product, including the ones it would not have caught: incidents.
  • Every control starts in dry_run and logs what it would have blocked before you switch it to enforce: trust.
  • Cost per agent and per owner (FinOps), priced per agent step rather than per user: pricing.

What this comparison does not claim

We haven't tested Agent 365. We make no claim about its detection depth, enforcement or cost reporting beyond what its product page says above.