All posts
Architecture Published Updated 11 min

November 10 agent guide: Gemini Robotics 2.5 and OpenShell

November 10, 2026 moved agentics into factories and Kubernetes. Google DeepMind released Gemini Robotics 2.5 for partner robots, and NVIDIA contributed OpenShell to the CNCF as a sandbox project for securing autonomous agents.

Jigar JoshiJigar JoshiAgentic AI Architect and Consultant
November 10, 2026 cover for Gemini Robotics 2.5 and OpenShell
In this post (7 sections)

November 10, 2026 is a physical-agent and agent-runtime day for Bacancy Technology readers. Google DeepMind introduced Gemini Robotics 2.5, its most capable vision-language-action model to date for partner robots. At KubeCon, NVIDIA contributed OpenShell to the CNCF Sandbox so organizations can govern autonomous agents with policy outside the model.

What shipped on November 10

November 10, 2026 primary developments
ReleaseSourceStatusBuilder action
Gemini Robotics 2.5Google DeepMindPartner rolloutTrack if your robotics OEM is in the partner set; do not assume public API
OpenShell → CNCF SandboxNVIDIA / CNCFOpen governance under ai-openshellEvaluate OpenShell or Cedar policies for agent sandboxes
Agent Substrate (related)Google-led / CNCFSandbox acceptance noted by NVIDIAWatch for long-running agent workloads on Kubernetes

Gemini Robotics 2.5

Gemini Robotics 2.5 is a vision-language-action model aimed at broader skills, faster learning, higher dexterity, and safer physical behavior. DeepMind evaluated more than 200 tasks, from tool use and deformable objects to delicate manipulation. Versus prior baselines, DeepMind reports 30% to 100% higher success rates by task type and 20% to 30% faster completion. The model works across robot types including Apptronik Apollo 2 and Franka Duo platforms.

Factory proof point and safety

  • Mercedes-Benz production pilots in Hungary and Germany used Apollo 2 for automotive kitting with a reported 96% success rate.
  • Inference-time reasoning guidance steers around physical hazards; DeepMind cites +33% safe completion around obstacles and +43% around nearby people.
  • ASIMOV-Jailbreak, a new open-source robotics jailbreak benchmark, is released; DeepMind reports Gemini 4 Argon at 99.5% jailbreak resistance on that suite versus 82% for a prior baseline.

Gemini 4 Argon also appears as a research accelerator: generating simulated practice from camera scenes, building digital twins, and helping diagnose failures. Bacancy Technology reads Robotics 2.5 as embodied agent progress for industrial partners, not a drop-in replacement for software coding agents.

OpenShell joins the CNCF

OpenShell manages and monitors agent behavior and enforces policy-based boundaries on what an agent can reach: files, networks, credentials, and inference endpoints. CNCF and NVIDIA emphasize enforcement in the infrastructure around the agent, not trust in the model alone. OpenShell is adopting Cedar as its policy language. The project moves toward the ai-openshell GitHub organization under CNCF Sandbox governance while NVIDIA continues core engineering.

Builder checklist

  1. 01
    Separate physical-agent and software-agent roadmaps
    Do not force Robotics 2.5 into a software backlog. Ask OEM partners for harness access and safety docs if physical automation is in scope.
  2. 02
    Map current agent privileges
    List file, network, secret, and model-endpoint access for each coding or browser agent before comparing OpenShell policies.
  3. 03
    Watch CNCF working groups
    Follow ai-openshell RFCs for observability, multi-tenancy, and Cedar policy patterns rather than forking an internal one-off sandbox.

Conclusion

November 10 connects factory-floor VLAs with Kubernetes-native agent runtimes. DeepMind’s Robotics 2.5 is partner-first physical intelligence. OpenShell’s CNCF move is the governance story software agent teams should track this week.

Need agent sandbox and policy design?

Bacancy Technology designs agent runtime controls, MCP allowlists, and production sandbox patterns. Consulting: /consulting/ . Contact: /contact/ .

Sources: deepmind.google ; blogs.nvidia.com — open source openshell infrastructure ; cncf.io .

The weekly take

Agentic AI patterns, delivered Thursdays

What I am shipping, watching, and pruning out of client stacks each week. One email. No fluff.

Shipping an agentic AI project this quarter?
Book a 30-min consult
Frequently asked

Questions readers ask about this post

Share this post
LinkedIn Facebook WhatsApp