Architecting efficient context-aware multi-agent framework for production- Google Developers Blog

What this post added

This post introduces the concept of 'context engineering' as a discipline for building production-grade AI agents, moving beyond simply increasing context window sizes. It details how the Google Agent Development Kit (ADK) implements this by treating context as a 'compiled view' over a richer stateful system. Key architectural principles include separating storage from presentation, using explicit transformations via processors in an LLM Flow, and enforcing scope by default. The post elaborates on the tiered model of context (Working context, Session, Memory, Artifacts), the role of Session Events for structured history, context compaction using LLMs to summarize older events, and context caching optimizations by treating context as stable prefixes and variable suffixes. The `contents` processor is highlighted for transforming Session Events into the working context for LLM calls.