Context management becomes a critical bottleneck when deploying production AI agents for long-running tasks. Google's Agent Development Kit (ADK) introduces a tiered context architecture that treats context as a compiled system rather than a simple string buffer. The framework separates storage from presentation through three
Table of contents
1.3 Session and events: structured, language-agnostic history1.4 Context compaction and filtering at the session layer1.5 Context caching2. Relevance: Agentic management of what matters now2.1 Artifacts: externalizing large state2.2 Memory: long-term knowledge, retrieved on demand3. Multi-agent context: who sees what, when3.1 Two multi-agent interaction patterns3.2 Scoped handoffs for agent transfer3.3 Translating conversations for agent transferConclusionSort: