Google Cloud Next '26: agents move from chat window to background workers

The shift from prompting to delegating

Google's headline framing is that AI has moved from transforming work to running at scale. The concrete evidence for that claim is the introduction of long-running agents that, per the announcement, 'work autonomously in the background within secure cloud sandboxes while you focus on other things.'

That is a different operating model than a chatbot you interact with turn by turn. A background agent working on a multi-step business process needs a place to run, boundaries on what it can touch, and a way for a human to check on it. Google addresses all three: the sandbox is the containment, and the new Agent Inbox is the supervision layer where users 'monitor, guide and manage' what agents are doing.

The Agent Inbox is the most telling detail. Google is implicitly acknowledging that once you have many agents acting on their own, the interface problem is no longer conversation — it's oversight. An inbox for machine work is a bet that people will manage fleets of agents the way they triage email.

Two audiences, two build paths

Google split the tooling by who is building. For technical teams, the Gemini Enterprise Agent Platform offers Agent Studio, a low-code interface, plus direct access to Gemini 3.1 Pro, Gemini 3.1 Flash Image (Nano Banana 2), Lyria 3, and — notably — Anthropic's Claude Opus 4.7. Offering a competitor's model inside its own platform is Google's stated 'open choice' position.

For everyone else, the Gemini Enterprise app ships a no-code Agent Designer for 'trigger-based workflows without writing a single line of code.' The repeated insistence that 'you don't have to be a machine-learning PhD' signals where Google thinks agent creation is headed: away from ML specialists and toward business users defining automations in natural language.

The risk in this framing is that governance is harder than authoring. Letting anyone build a trigger-based agent expands who can automate a process; it does not automatically expand who understands the failure modes. The Agent Inbox and sandboxes are the mitigations Google is offering, but the burden of defining safe triggers still lands on non-specialists.

Inference economics baked into silicon

The infrastructure announcement is more specific than usual. Google split its eighth-generation TPUs into two roles: the TPU 8t for training and the TPU 8i for inference, with the inference chip cited at '80% better performance per dollar.'

That split matters for anyone planning to run agents at volume. Training is a one-time-ish cost; inference is what you pay every time an agent acts. Optimizing a dedicated chip for serving — and quoting the metric in performance per dollar rather than raw speed — reflects Google's own argument that 'running millions of AI agents takes some serious computing muscle.' The cost of autonomous work, not just its capability, is the constraint being engineered against.

The supporting pieces reinforce data movement as the bottleneck: the Virgo Network to connect supercomputers, and Managed Lustre cited at 10 terabytes per second. Google is also positioning as an early host of NVIDIA's Vera Rubin NVL72 alongside its own Axion processors, keeping a multi-vendor compute story.

Letting data stay put is the quieter, bigger claim

The Agentic Data Cloud carries the line most relevant to real deployments: an agent 'is only as helpful as the information it understands.' Two components back this. The Knowledge Catalog uses Gemini to autonomously tag and connect enterprise data so agents grasp company-specific context. The Cross-Cloud Lakehouse, standardized on Apache Iceberg, lets you query data where it already lives.

Standardized on Apache Iceberg, this lets you leave your data exactly where it is — even if it's in AWS — and query it instantly, with zero friction.Montana Labs

Explicitly naming AWS as a place your data can stay is a pointed move. It concedes that customers will not migrate everything to run Google's agents, and it lowers the switching cost of trying them. For automation specifically, this addresses the practical blocker: agents fail when they lack context, and context usually sits scattered across systems no one wants to consolidate first.

The named deployments — Home Depot's in-store assistant, Papa John's Ordering Agent that remembers 'the usual,' Mars and Citadel Securities on quantitative research, and Unilever deploying agents across an organization serving 3.7 billion consumers — describe agents plugged into existing operations rather than greenfield builds. That is the coherent thread across the seven highlights: Google is optimizing for automation that runs on top of the messy data and infrastructure companies already have, not automation that requires rebuilding first.