Agent Runtimes Hit the Metal: Keyless microVMs and a Concurrency Engine

Highlights signal board

• signals03
• sources05
• watch02

• Google Cloud unveiled Agent Compute Engine, giving each agent its own hardware-isolated microVM with no credentials stored inside - the infra it already runs Gemini on. (The New Stack)
• AWS open-sourced Infinity, an Apache-2.0 Rust runtime for "extreme agent concurrency" that suspends and resumes agents around tool calls, and introduces RAP, an async successor to MCP. (AWS Open Source)
• OpenAI's Decisions API went generally available at $0.10 per million input tokens (output free), returning typed, bounded answers for routing and classification. (TechCrunch)
• The Agentic AI Foundation added Open Responses, OpenEnv, and OpenSandbox, reaching 10 hosted projects in its first year. (AAIF)

Key Signals ranked scan

1. Google turns its Gemini agent infra into a cloud service Oct 22, AGNTCon + MCPCon

   Agent Compute Engine gives every agent a persistent, sandboxed microVM with root access that can run shell, a headless browser, and code "in loops that could run for weeks," then checkpoints and pauses between tool calls. Long-lived secrets never live inside the VM - they're injected at egress by an Agent Gateway only after mTLS and policy checks. Preview opens "in the coming months," no pricing yet. (The New Stack)

   runtime

2. AWS reimagines the runtime for thousands of idle agents Oct 23, Apache 2.0

   Built with UC Berkeley under the Hydro project, Infinity dispatches a tool call, saves conversation state, and suspends the agent until the result arrives - borrowing async suspend/resume from programming-language runtimes. Existing stdio/HTTP MCP servers run unchanged through a compatibility layer, while the new Reactive Agent Protocol (RAP) drops long-lived connections entirely. (AWS Open Source)

   concurrency

3. A dedicated endpoint for machine decisions, not chat GA Oct 22

   The POST /v1/decisions endpoint, powered by GPT-6 Luna, returns a probability, a selection, or a rubric score - roughly 10x faster than the Responses API at ~150 ms. At $0.10/M input and free output, it reframes classification, routing, and prioritization as a cheap, typed call rather than a prompt-parse loop. (TechCrunch)

   model ops

Why It Matters / What To Watch watch items

1. The runtime layer is becoming cloud infrastructure, not an SDK.
   • Yesterday's story was SDK-level hardening; this week the hyperscalers moved it down the stack. Watch whether ACE's keyless, egress-injected credential model becomes the default pattern for long-running agents. (The New Stack)
   • Track RAP adoption versus MCP: Infinity keeps MCP working but bets the future is async. If concurrency and cost (75k idle agents per host, AWS claims) hold up, long-horizon agent fleets get much cheaper to park. (AWS Open Source)
2. Standards and cross-vendor governance keep consolidating.
   • AAIF now stewards 10 projects under neutral governance; Open Responses in particular promises API portability across providers - worth watching if you're locked to one SDK. (AAIF)
   • Dataiku Agent Management reached GA this month, inventorying agents across AWS Bedrock, Databricks, Google Vertex, Microsoft Copilot Studio/Azure Foundry, Salesforce Agentforce, and Snowflake Cortex. As runtimes multiply, "which agents are even running, and what risk do they carry" becomes the operator's first question. (Dataiku)

Quick Links sources

• Google turns the infrastructure behind Gemini's agents into a cloud service The New Stack
• Introducing Infinity: a runtime for extreme agent concurrency AWS Open Source Blog
• OpenAI makes its Jev competitor widely available as decision model competition heats up TechCrunch
• Agentic AI Foundation Announces Four New Projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox Agentic AI Foundation
• Dataiku Launches Agent Management for enterprise AI agents Dataiku