Chronological feed of everything captured from Sam Charrington.
Google used Cloud Next '25 to deliver a meaningful overhaul of its AI agent platform, replacing the failed Dialogflow-based Agent Builder with a code-first suite: the Agent Development Kit (ADK), Agent Engine (managed runtime), and Agent Garden (agent repository). This addresses the "product debt" the author flagged at Next '24, when the original Agent Builder proved too rigid for real-world agentic workflows. However, while foundational agent tooling is now more credible, Google simultaneously launched new immature offerings — Agentspace, A2A protocol, Agent Engine — that reset the debt clock. Gemini models now dominate the cost-performance Pareto frontier across LLM benchmarks, giving Google a strong underlying position even as platform complexity and product identity churn remain persistent risks.
Google Cloud Next '26 marked a shift from architectural vision to operational delivery: Vertex AI is being retired in favor of the unified Gemini Enterprise brand, and the platform's center of gravity has moved into production-grade agent operations—identity, governance, observability, simulation, and evaluation. The new Agent Platform is organized around four verbs (Build, Scale, Govern, Optimize), with the Govern and Optimize layers receiving the bulk of net-new investment. Google also split its 8th-gen TPU line into dedicated training (TPU 8t) and inference (TPU 8i) chips—a long-cycle hardware bet that signals inference is now the strategic workload. Critically, most announcements shipped in Preview or GA at the event itself, a meaningful departure from prior years' "coming soon" pattern.