2026-W38 · 12 stories

The week in AI: September 14–20, 2026

The week's main developments, grouped by original publication date. Every item links to our full story and the original source.

Issue date: ·Week: September 14–20, 2026

What changed this week

The selected stories cover new voice-agent models, AWS agent execution, and Google Cloud infrastructure security changes during the reporting period.

Three key changes

  1. Google introduces Gemini 3.8 Live for voice agents

    What changed
    Google introduced Gemini 3.8 Live and an Extended Thinking variant for real-time voice interaction and conversation-based task completion.
    Practical impact
    Developers can plan for real-time voice tasks that also complete actions, using the model alongside existing agent flows.
    What to consider
    Article does not state dialogue runs or measurable performance against a specific call flow; no release calendar for the September 2026 update is provided, so rollout timing is unverified.
    Full story ↗
  2. AWS Launches AgentCore Runtime for Consistent Agent Start Times

    What changed
    AWS introduced AgentCore runtime with snapshot-based startup optimization and dynamic memory reclamation for more predictable AI agent deployment.
    Practical impact
    Teams can expect more consistent cold-start behavior, with dynamic memory reclaimed during sessions instead of holding memory until session end.
    What to consider
    The examples use 200 MB and 2 GB images, but no independent benchmark data or public pricing is stated for the new runtime.
    Full story ↗
  3. Google deploys AI agents to secure infrastructure and automate code patching

    What changed
    Google uses its Mantis multi-agent framework for pre-submit scanning, reachability validation, and automated code patch generation, with claims of hundreds of issues avoided and over 92% precision in under a minute.
    Practical impact
    Code repositories can shift security checks earlier and reduce manual triage, while exact defect counts should be treated as vendor-reported.
    What to consider
    Google reports thousands of vulnerabilities and pre-submit scanning issues, but independent verification of the security impact and production adoption is not provided.
    Full story ↗

Models

1

Google introduces Gemini 3.8 Live for voice agents

Google's new Gemini voice models combine live conversation with task completion. What developers should know.

This could help developers build applications where users speak a request and expect an action, not just an answer. Performance and limitations still need to be assessed against a specific conversation flow.

Full story ↗Original source ↗

Products

2

Google Cloud Launches Filestore Agent Volumes for AI Workloads

Google Cloud introduces Filestore agent volumes, a managed storage service designed for elastic and isolated AI agent workspaces in GKE.

By automating the provisioning of isolated persistent storage, the service reduces storage-related cold-start latency and infrastructure overhead. It addresses the complexity of managing file systems for dynamic, ephemeral AI agent deployments.

Full story ↗Original source ↗

Salesforce unveils AIforce to bring CRM data into Claude and Slack

Salesforce announced AIforce at Dreamforce, launching with Claudeforce, Slackforce and Agentforce Coworker. Salesforce in Claude is in beta. Pricing may change; availability varies by region and customer agreement.

Companies on Salesforce could let teams work with CRM data from Claude and Slack without a new permissions model or custom integration, if Salesforce's claims hold. Those claims come from the vendor and have not been independently tested.

Full story ↗Original source ↗

Applications

2

Siemens Uses Agentforce for Leads and Plans Teamcenter Integration

Two Salesforce agents handle Siemens' inbound leads; the companies have announced a separate Teamcenter integration for sales and service.

The case distinguishes a live lead-handling workflow from a proposed engineering-data integration. Siemens says the two agents now engage every inbound lead; the effects of connecting Teamcenter to sales and service remain to be measured.

Full story ↗Original source ↗

Developers

7

Agent Substrate is available on GKE for non-production workloads

Google Cloud has made Agent Substrate, an open-source agent runtime, available on GKE. Production support requires allowlist access, and the performance figures are company claims.

For teams running many isolated agents, the key questions are how densely sandboxes can run and how well they are isolated. Optional Filestore agent volumes add shared NFS storage with RWX access and POSIX file locking. The announcement does not state pricing.

Full story ↗Original source ↗

Salesforce and Google Cloud Unify Infrastructure and AI Agents

Salesforce and Google Cloud have expanded their partnership to integrate Salesforce's headless architecture with Gemini Enterprise and enable native interoperability.

This integration enables enterprises to deploy autonomous AI agents that process data across Salesforce and Google Cloud ecosystems without the need for custom connective software.

Full story ↗Original source ↗