Give Your Agents Tools That Scale Down to Zero or Up to Infinity

Originally shared by Google Cloud Tech on X - August 26, 2026. By @zeroasterisk and @Saboo_Shubham_.

An agent is only as capable as its tools, and in production those tools run as Model Context Protocol (MCP) servers. Every other layer of your stack scales to zero when it goes idle — until last month, a single MCP session pinned every tool call to one container.

Key Points

• Sessions were an unadvertised tax: round-robin load balancing broke older, stateful MCP servers with 400 Session Not Found errors.
• The 2026-07-28 spec makes every tool call self-describing: SEP-2575 and SEP-2567 delete the initialize handshake and the Mcp-Session-Id header, moving identity into inline _meta.
• Tool containers can finally scale to zero: deploy stateless MCP tools to Cloud Run with --min-instances=0 and retire the Redis session cluster.
• Wiring stateless tools into ADK takes six lines of Python: point McpToolset at a StreamableHTTPConnectionParams URL — no session choreography required.
• Stateless is the transport, not your application: long-running work gets an explicit task handle via the Tasks extension instead of hiding state in a socket.

Why This Matters

This is the companion field report to my Google Developers Blog post on scaling AI agent infrastructure — the same story, with the exact curl reproduction steps, the ADK wiring code, and the Cloud Run deploy command. SDK betas for the 2026-07-28 spec are live for Go (v1.7.0), Python (mcp[cli]==2.0.0b1), and TypeScript.

View the original thread: Google Cloud Tech on X