LLMs are inherently stateless, making agentic reliability highly dependent on context engineering. Even a 5% drop in per-step accuracy can cause a 14-point drop in reliability over three chained agent calls. Getting context right is the difference between a demo and a production system.

This session covers how to optimize the context window with tools, skills, memories, and history using ADK. We will explore on-demand tool retrieval, progressive-disclosure skills, and context compaction for long-running agents. We will dive deep into Memory Bank to see how noisy session events become refined, durable memories, how to customize them via topic definitions and few-shot examples, and when to inject them.

Leave with a framework for context control, state persistence, action routing, and insights into the 2026 Gemini Enterprise roadmap.