KDCube
← AI Industry News

Agent Runtimes Learn to Sleep: AWS, Google Snapshots, and Obsidian's Policy Gate

AWS and Google both shipped snapshot-and-restore agent runtimes within 48 hours — Bedrock AgentCore cuts cold starts to under two seconds while Google's open-source Agent Substrate heads to the CNCF with a claimed 10x cost cut. Meanwhile Obsidian Security moved governance to the action layer, and banks...

Lead signals

Highlights

  • AWS shipped a next-gen Bedrock AgentCore Runtime that snapshot-restores sessions in under two seconds, mounts S3 via FUSE, and scales idle agents to zero — Coinbase says cold starts dropped "from nearly a minute to under five seconds." (unite.ai)
  • At KubeCon NA, Google donated Agent Substrate (plus OpenRL and KubeRay) to the CNCF, pitching suspend-and-snapshot scheduling for agents that sit idle 90–95% of the time and claiming a 10x cost cut versus dedicated VMs. (SiliconANGLE)
  • Obsidian Security launched a Runtime Policy Decision Engine that evaluates every agent action — across Claude Code, Codex, Copilot, Bedrock, and Vertex — before execution, with an MCP proxy that keeps app tokens out of the AI client. (Help Net Security)
  • The counter-trend: banks, retailers, and airlines are actively blocking AI agents from their systems, a reminder that runtime economics are racing ahead of trust. (AI Agents Directory)

Key Signals

  1. Snapshot-restore becomes the default runtime primitive Nov 12–13, 2026

    Within 48 hours, both AWS and Google shipped the same core idea: freeze an idle agent's memory and files, free the compute, and restore on demand. AgentCore v2 now offers 8 vCPU / 32 GB sessions, 10 GB images, and committed-baseline pricing, with customers reporting up to 68% less memory per session and ~60% smaller footprints (Zuora). Google's Agent Substrate makes the same bet open-source on Kubernetes. The economic insight is identical — agents spend most of their life waiting on models, tests, or human approval — and operators now have two production-grade ways to stop paying for that idle time. (unite.ai; SiliconANGLE)

  2. Governance moves to the action layer, not just the credential Nov 11, 2026

    Obsidian's engine enforces allow/block/human-approval decisions using identity, employment, app access, and data-classification context, and extends to customer gateways via a Policy Decision API. CEO Hasan Imam framed the gap bluntly: "An agent can hold valid credentials, use an approved tool and still take an action that costs the business millions." For teams running many agents, the control point is shifting from who the agent is to what the action does. (Help Net Security)

  3. Trust lags the tooling Nov 12, 2026

    Even as runtimes get cheaper, PCWorld reports enterprises blocking agent traffic outright, and Palo Alto integrated Cortex Cloud with Claude Code and Cursor to secure AI-generated code — citing that roughly 45% of such code introduces vulnerabilities. The infrastructure is maturing faster than the guardrails around it. (AI Agents Directory)

Why It Matters / What To Watch

  1. Re-cost your long-running agents now.
    • If you run multi-turn or always-on agents on AgentCore, test v2's scale-to-zero and committed-baseline billing against your current spend — the 60–68% memory reductions are per-session, so savings compound with concurrency. (unite.ai)
    • If you're self-hosting on Kubernetes, watch Agent Substrate's CNCF trajectory before building your own idle-reclaim scheduler. (SiliconANGLE)
  2. Put a policy gate between agents and your systems.
    • Evaluate runtime action-approval (Obsidian-style MCP proxies) so agents can't take a correct-looking but catastrophic action with valid credentials. (Help Net Security)
    • Expect more customers and partners to block unauthenticated agent traffic; design for explicit, auditable agent identity now rather than retrofitting it after a block. (AI Agents Directory)

Quick Links