Google Cloud Next '26 Developer Keynote

The Google Cloud Next '26 Developer Keynote, running approximately 67 minutes, showcases Google's Gemini Enterprise Agent Platform through a live end-to-end demonstration of building a multi-agent marathon planning and simulation system for Las Vegas. The keynote is hosted by Richard Seroter and Emily (Emma) Twersky, with a series of Google engineers and a guest from Wiz delivering focused demos on specific platform capabilities.

Brad Calder opens by introducing the Gemini Enterprise Agent Platform's core components: the Agent Development Kit (ADK) for building agents, scalable serverless agent runtime, session and memory management, agent identity and governance via Agent Gateway, Agent Registry for discovery, the A2A (Agent-to-Agent) protocol donated to the Linux Foundation, and agent observability with integrated eval.

Mofi Rahman demonstrates building a planner agent using Agent Designer (low/no-code) and then extending it with code. The agent is given skills (YAML + Markdown bodies) covering mapping, GIS, and race director expertise, plus tools via Google Cloud's fully managed MCP servers for Google Maps. Every Google Cloud service is now MCP-enabled by default.

Casey West and Ivan Nardini show multi-agent orchestration: a planner agent spawns an evaluator sub-agent (using a separate model with limited context) and a simulator agent. The A2A protocol eliminates brittle API contracts between agents through agent cards. The Agent Registry acts as a DNS for agents — deployed agents are automatically registered. The demo introduces ATUI (Agent-to-User Interface), an open standard that lets agents dynamically generate their own UI components, moving beyond walls of text.

Lucia Subatin and Jack Wotherspoon cover context engineering and memory. They show fewer than 20 lines of ADK code enabling session management and memory bank integration. AlloyDB's auto-embeddings feature generates vector embeddings automatically. A data engineering agent builds RAG pipelines from a natural language prompt using Lightning Engine for Apache Spark and Document AI for semantic chunking. The planner agent pulls long-term memories from past simulation runs to improve future route planning.

Megan O'Keefe demonstrates agent observability and debugging at scale. She uses agent runtime trace views to identify a cascading 400 error caused by exceeding Gemini's 1 million context token limit. Cloud Assist (an AI operator agent connected via MCP) investigates logs, searches GitHub for relevant issues, pinpoints the root cause in source code, and suggests a proactive fix — adding a token threshold to event compaction config — which is committed and deployed via CI/CD.

Bobby Allen shows 'vibe clouding': using Gemini Cloud Assist to convert an existing Cloud Run service to GKE and deploy a fine-tuned Gemma 4 model, all from a natural language prompt. Cloud Assist proactively identifies a storage bottleneck (GCS Fuse too slow at scale) and recommends switching to Lustre. The runners in the simulation now display Gemma 4-generated thought bubbles.

Jason Davenport and Ines Envid demonstrate no-code agent creation in Gemini Enterprise. A marathon supply chain agent team (main agent + sub-agents for specific domains) is created from a single prompt with a Google Doc as context. This agent works alongside the high-code planner agent via the shared Agent Registry.

Ankur Kotwal covers agent security: Agent Gateway enforces IAM policies as a proxy between agents, agent identities provide unique immutable credentials (compared unfavorably to shared service accounts), and egress policies prevent agents from accessing unauthorized resources. A live demo shows a read-only finance policy blocking the planner agent from writing budget changes.

Yinon Costica from Wiz demonstrates code security integration: Wiz's security graph maps the full agent architecture and attack paths. A red agent (penetration tester) validates exploitable vulnerabilities including an authentication bypass. A green agent suggests three prioritized fixes, which a coding agent applies directly in the IDE and commits, triggering a Wiz rescan that confirms the fixes.

The entire simulation codebase — including all agents, the Angular/Flutter UI, and infrastructure — is open-sourced and published to GitHub with a developer solution guide and free cloud credits for labs.