The Week in Cloud & AI: GKE Suspends Idle Agents and Restores Them in Under a Second

The Week in Cloud & AI

Agents stopped being a model feature this week and became a workload. Google gave GKE a runtime that suspends idle agents and restores them in under a second, AWS raised Lambda's asynchronous timeout sixfold to 90 minutes, and OpenAI shipped a managed Agents API. Running alongside that, four vendors independently shipped controls over what agents may read and what they may spend. The infrastructure is catching up with something teams have already put into production.

ReleaseBytes Insights

Two halves of the same problem landed in the same week, from vendors with no reason to coordinate.

The first is scheduling. Agents are long-running, mostly idle and bursty, which is the worst possible shape for infrastructure built around short synchronous requests. GKE's Agent Substrate snapshots an idle agent's memory and local files and restores it in under a second, deliberately bypassing Kubernetes control plane bottlenecks to fit more agents per machine. AWS attacked the same mismatch from the other end, lifting Lambda's asynchronous ceiling from 15 to 90 minutes. Both are admissions that the existing primitives did not fit.

The second is containment. GitHub now lets administrators centrally restrict which shell commands, files and network domains an agent may touch, with no user override. Databricks is narrowing Genie agents to explicitly attached data sources. Snowflake and Microsoft approach the same question through data classification and cost.

Four vendors converging on agent permissions in one week is strong evidence, not an emerging signal. Teams choosing an agent platform over the next two quarters should weigh its permission model as heavily as its model quality.

If You Only Read One Thing...

GKE's Agent Substrate is the week's most structurally interesting release. It treats an idle agent as something to suspend and restore rather than keep running, which is the first credible answer to what agent workloads cost at scale. If your roadmap assumes a pod per agent, the economics just changed underneath it.

In This Edition

• Top Stories
  • GKE Adds a Runtime That Suspends and Restores Idle Agents
  • AWS Lambda Raises Asynchronous Timeouts to 90 Minutes
  • OpenAI Ships a Managed Agents API
  • GitHub Puts Agent Permissions Under Central Enterprise Control
• Breaking Changes
• Security Updates
• Attention Required
• Industry Trends

Top Stories

GKE Adds a Runtime That Suspends and Restores Idle Agents

Google Kubernetes Engine gained Agent Substrate, for running agentic workloads at scale. It suspends idle agents, snapshots their working memory and local files, and restores that state onto an available sandbox with sub-second latency when the agent is next triggered. Google says it bypasses standard Kubernetes control plane bottlenecks to support more concurrent agents per machine. It is available for evaluation and non-production use, with production support gated behind a limited GA allowlist.

Why it matters The naive way to run agents is a long-lived pod each, billing for hours of idleness. Substrate makes idle cost near zero without losing session state, and moves the scaling limit off the control plane. The allowlist gate makes this a direction of travel rather than something to build on this quarter.

Key Takeaways

• Suspends idle agents, snapshotting memory and local files.
• Sub-second restore onto an available sandbox.
• Evaluation and non-production today; production via allowlist.

Who should care? Platform Engineers, AI Engineers, Technical Architects Impact High

Read the full summary on ReleaseBytes

AWS Lambda Raises Asynchronous Timeouts to 90 Minutes

AWS increased the function timeout for asynchronous and event source mapping invocations on Lambda Managed Instances from 15 to 90 minutes. It covers data processing, media transcoding, financial calculation, AI inference and batch workloads, including inside Lambda durable functions. Synchronous invocations keep the 15-minute maximum.

Why it matters The 15-minute ceiling has been the standard reason to abandon Lambda for Step Functions, Fargate or a queue-and-worker design. Removing it on asynchronous paths deletes a class of architecture written purely to work around it. The synchronous carve-out matters, so check the invocation type before assuming this applies.

Key Takeaways

• 15 to 90 minutes for asynchronous and ESM invocations on Managed Instances.
• Synchronous invocations still capped at 15 minutes.
• Removes a common reason to re-architect onto Step Functions or Fargate.

Who should care? Platform Engineers, DevOps Engineers, Software Engineers Impact High

Read the full summary on ReleaseBytes

OpenAI Ships a Managed Agents API

OpenAI introduced the Agents API, a managed service for building and running cloud agents, built on the Codex harness. It offers orchestration, long-running sessions and tool use as platform features rather than things each team assembles. GPT-Live-1 arrived on the same API, adding full-duplex voice with custom voices and direct telephony.

Why it matters Most production agent stacks are bespoke: a queue, a state store, retry logic and a tool registry written in-house. Moving orchestration and session persistence behind a vendor API removes that undifferentiated work, at the cost of coupling the control loop to one provider. That trade is the central architectural decision for agent teams this year.

Key Takeaways

• Managed orchestration, long-running sessions and tool use.
• GPT-Live-1 adds full-duplex voice and telephony to the same API.
• Shifts agent control-loop code from application to platform.

Who should care? AI Engineers, Software Engineers, Technical Architects Impact High

Read the full summary on ReleaseBytes

GitHub Puts Agent Permissions Under Central Enterprise Control

GitHub Copilot Business and Enterprise administrators can now centrally control what agents may do: which shell commands run, which files are readable, which network domains are reachable. The restrictions cannot be overridden by user or workspace settings, and separate policies can target different teams. Generally available across the Copilot app, CLI and IDE integrations.

Why it matters Until now the realistic choice for a security team uneasy about agent shell access was to disable agent workflows entirely. Central, non-overridable policy replaces that with something proportionate, and makes agent use defensible in regulated environments. The unbypassable part is what separates this from a setting.

Key Takeaways

• Central control over shell commands, file access and network domains.
• Cannot be overridden by user or workspace settings.
• Per-team policies, generally available across app, CLI and IDE.

Who should care? Security Engineers, Engineering Managers, Platform Engineers Impact High

Read the full summary on ReleaseBytes

Breaking Changes

• Google Cloud SDK 584.0.0 ships breaking changes alongside deprecations and a security update. Review before upgrading pinned CI images. Details
• Databricks Java SDK v0.153.0 introduces API changes that break compilation. Pin the previous version until call sites are migrated. Details
• Cloud SQL for PostgreSQL and MySQL temporarily disabled sqlcommenter tag appending on remote MCP. Query attribution built on those tags will show gaps.

Security Updates

• npm extended its 72-hour security hold to all accounts. After a recovery-code sign-in, publishing and token creation pause for three days. Automatic, but plan releases around it if you rely on recovery codes. Details
• Google patched Slurm vulnerabilities in Cluster Toolkit, including an sbcast issue. HPC and training clusters should take v1.103.0. Details

Attention Required

• Databricks Genie agents lose implicit data access in late September 2026. Agents reach only sources explicitly attached under Sources, not tables merely referenced in instructions. Audit agent definitions now. Details
• Google SecOps removes write permissions from the chronicle.readonly OAuth scope on 25 January 2027. Automation writing through that scope must move to chronicle. Details
• Amazon Bedrock retired anthropic.claude-3-haiku-20240307-v1:0. Pinned inference profiles referencing it now fail. Details
• GitHub Copilot is deprecating MAI-Code-1-Flash. Update pinned model configuration. Details

Industry Trends

Agent permissions are becoming a product category. GitHub shipped central agent policy, Databricks narrowed Genie's data reach, Snowflake previewed intent-driven governance turning plain-English rules into masking policies, and Microsoft published agent cost governance with token rate limits and project-level cost tagging. Four vendors, one week, no coordination.

Voice is moving from feature to interface. GPT-Live-1 brings full-duplex conversation and telephony to the OpenAI API, putting agents on phone lines without a separate media stack.

Idle cost is the new scaling constraint. GKE's snapshotting and Lambda's longer timeouts target the same economics: workloads that wait far longer than they compute.

By the Numbers

• 304 releases analysed
• 9 breaking changes
• 12 security updates
• 43 deprecation notices
• 149 GA releases

Thanks for reading. Browse every announcement from the week at ReleaseBytes, and subscribe to get this digest each Monday. If a colleague is tracking agent infrastructure, forward this one on.