Google's Agent Compute Engine: checkpointed microVMs stop paying for agent idle time
Arindam_1729 · x · 2026-11-02
Google Cloud's deep dive on Agent Compute Engine (ACE) tackles a core coding-agent problem: agents spend most of their life waiting on the LLM, and traditional options (rebuild the sandbox each step, or keep a VM alive all session) don't scale.
Architecture: agents split into an Agent Harness (calls the model, decides next action) and an Execution Sandbox — a hardware-isolated microVM running code, shell, and browser actions.
Three key design choices:
- Checkpoint & restore: idle microVMs are snapshotted and restored in place when the next action arrives — no rebuild, no idle cost
- Copy-on-write memory: unchanged pages of a shared OS image and Python runtime are shared across microVMs, so the same hardware hosts far more agents
- Isolation & persistence: each session gets its own microVM with a guest kernel; workspaces live on Hyperdisk across sessions; credentials stay outside the sandbox behind external gateways
Tradeoff: restore latency sits in the agent loop, so it only wins when idle gaps exceed restore time — true for most coding agents. Lesson: treat compute like a pausable process, not a server you keep running.
Related event: Google Unveils Agent Compute Engine for Billion-Scale Agents(2 posts)→
More from coding & agent
- Overnight coding agents stall on questions: bidirectional answer routing is the missing piece — onitsoga · 2026-11-02
- Testing AI models via game-ready 3D assets: mix Blender with Tripo and Meshy — chongdashu · 2026-11-02
- Bivia by Articul8 skips the paragraph: a decision model that returns JSON with confidence and evidence — No-Asparagus-8011 · 2026-11-02
- dhh: agent era ends tech talent's immunity for being difficult, soft skills matter — bendee983 · 2026-11-02
- How to edit videos with Claude Code and Opus 5.5, plus 32 shared skills — aakashgupta · 2026-11-02
- Claude Code can edit video — the trick is dictating editor notes while recording — aakashgupta · 2026-11-02