Google's SAM is a zero-trust P2P mesh for AI agents

What it does

SAM is a networking layer for autonomous AI agents. It runs lightweight sam-node clients that form a self-healing P2P mesh using libp2p, allowing agents to plug in and dynamically discover and invoke tools across the network. A central control plane handles identity registration and authorization policies, while routers manage data-plane connectivity. The goal is to let agents communicate and share capabilities as if they were on the same local mesh, regardless of whether they run in the cloud, on a laptop, or at the edge.

The interesting bit

The project treats agent-to-agent networking as a zero-trust infrastructure problem rather than an afterthought. Every node, connection, and packet is cryptographically authenticated, and identities are portable across environments. It also sidecars standard MCP (Model Context Protocol) traffic onto the mesh, so agents built for Gemini, Claude, or similar frameworks can use remote tools without knowing the network topology.

Key highlights

• Nodes auto-discover peers and build the mesh with zero manual configuration.
• Strict zero-trust posture: all connections, nodes, and packets are authenticated.
• Agents use standard MCP to call tools hosted by other nodes across the mesh.
• Cryptographic identities are environment-agnostic, enabling node mobility between cloud, local, and edge.
• Architecture splits responsibilities into a control plane (sam-control-plane), libp2p routers/relays (sam-router), and local node clients (sam-node).

Caveats

• Google explicitly states this is not an officially supported product.
• Running your own hub requires deploying the full control plane, router, and console stack.

Verdict

Worth a look if you’re building multi-agent systems that need to share tools across environments without exposing public APIs. Skip it if you just need a single agent to call a static set of local functions.