AI Agent Python SDK: Google ADK (google.github.io)

Google’s Agent Development Kit (ADK) introduces first-class support for multi-agent systems (MAS) in its Python SDK, letting developers decompose complex agentic applications into composable BaseAgent instances. You can combine LLM-powered agents, workflow orchestrators (SequentialAgent, ParallelAgent, LoopAgent), and custom non‑LLM agents into hierarchical trees by passing sub_agents at construction; ADK automatically sets parent_agent and enforces a single-parent rule to keep ownership clear. This structure boosts modularity, specialization, reusability and makes it easier to reason about control flow and delegation in large agentic applications. Technically, ADK provides primitives for orchestration and intra-system communication: SequentialAgent forwards the same InvocationContext so downstream agents can read state written by upstream steps (via output_key), ParallelAgent runs children concurrently while isolating history using InvocationContext.branch but still shares session.state (so distinct keys avoid races), and LoopAgent repeats sub-agents until max_iterations or an Event with actions.escalate=True terminates the loop. Agents communicate primarily through session.state and the InvocationContext.temp for per-turn data; LLM agents’ outputs can be auto-saved to state keys. These features enable reliable pipelines, parallel data collection, and polling/monitoring patterns—making complex LLM workflows more maintainable, debuggable, and scalable.