Teleport launches Beams, giving each AI agent its own microVM and identity

Teleport has announced Beams, a trusted runtime for AI agents that operate against production infrastructure, with an MVP dated 30 April 2026. Each Beam runs an agent inside an isolated Firecracker microVM with identity built in, connected to infrastructure and inference services without secrets and wired into Teleport's access control and audit trail. Teleport's framing is that teams currently stitch together IAM, infrastructure, and secrets by hand for every agent, with no consistent identity, no visibility into what an agent did, and each team reinventing its own container or VM workflow. Target uses run from internal agents reaching production services to multi-agent pipelines that need reproducible isolation.

Why it matters

Most agent-identity products this year issue a credential and govern what it can call. Beams starts a layer lower: the runtime the agent executes in is itself the identity boundary, so the blast radius of a prompt injection is one short-lived microVM rather than a long-lived container holding a broad token.

That is closer to how workload identity already works for services, and it is a better fit for infrastructure agents specifically, where the failure is not a bad API call but a shell command against production.

The trade-off is scope. This governs agents you deploy on infrastructure you run, not SaaS-hosted assistants or third-party MCP servers, so it is a complement to a broker rather than a replacement. It also deepens a Teleport dependency in a category where standalone vendors keep getting acquired: see the acquisitions tracker.

Source: Teleport