Google Cloud unveils Agent Compute Engine, the microVM platform behind Gemini agents: one hardware-isolated computer per agent, paused while it thinks, with no credentials inside; preview in the coming months
0 reads
- Published
- 2026-10-22
- Type
- Product
- Organization
- Source
- agipulse
Why it matters
【Key points】
On Oct 22 at AGNTCon + MCPCon, Google Cloud announced Agent Compute Engine: a hardware-isolated microVM per AI agent. It already runs Gemini agents, Antigravity and Google's eval/RL jobs; customer preview is due "in the coming months".
1. Pause while thinking: when a tool call ends, the host snapshots and idles the microVM, then resumes it on the next call. Agents share one pre-warmed image in memory (copy-on-write).
2. No secrets inside: own kernel and root, but no normal NIC, no metadata server, no tokens or keys. Agent Gateway checks user delegation, a short-lived SPIFFE agent ID and node identity, then swaps in real credentials off-VM.
3. A Hyperdisk per microVM keeps workspaces across days; RL/eval jobs can fork one snapshot into thousands.
4. Governance: Gateway, Registry and Identity will extend to other clouds, on-prem, and local coding agents like Claude Code and Antigravity CLI ("preview soon"). Per-agent token budgets and model routing preview this quarter.
【Verifiable】
https://cloud.google.com/blog/topics/ai-infrastructure/introducing-agent-compute-engine-and-expanded-agent-governance/
https://cloud.google.com/blog/topics/ai-infrastructure/agent-compute-engine-deep-dive/
https://cloud.google.com/blog/products/containers-kubernetes/agent-substrate-available-on-gke
https://agipulse.org/e/1014
https://agipulse.org/e/910
【Our read】
1. Agent runtimes are becoming a cloud category. Perplexity open-sourced SPACE, NVIDIA shipped OpenShell, Microsoft MXC, Walmart an MCP gateway, all within a month.
2. The design is close to SPACE. Google's edge is silicon root of trust plus identity wired into Workspace and VPC data, and app-aware idling that pushes idle-agent cost toward zero.
3. Keeping credentials out of the sandbox limits prompt-injection damage. Compare the LiteLLM token leak.
4. Caveats: no price or date. The 200k-agent, 500 activations/s, <500 ms and 10x density figures come from GKE Agent Substrate, not ACE.
Related developments
Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy