# Google Agent Development Kit (ADK) Google's framework for building AI agents, available in Python (since March 2026) and Go. Version 2.0 is a response to a production reality: agents that orchestrate everything through the LLM prototype well and then fall apart in production, with infinite loops, hallucinated routing, and runaway costs. The 2.0 answer is a structured workflow runtime that separates execution control from language processing. You compose a directed graph where: - **Deterministic nodes** handle routing, scheduling, and data transformation in traditional code - **Agent nodes** handle the ambiguous, unstructured inputs that genuinely need reasoning - **Programmatic routing** replaces LLM-based path decisions Google calls the resulting hybrids "Agentic Workflows", and the design principle behind them deserves its own note: [[Use code for control flow, agents for reasoning]]. Their guidance for choosing is refreshingly plain: predefined business logic goes in workflow steps; subjective, unstructured tasks go to agents. The claimed payoffs are concrete: context isolation (only necessary data passes between nodes, so no context bloat), roughly 50% token savings and 20% latency improvement over pure-agent approaches, and prompt injection mitigation, since fixed graph boundaries mean injected instructions can't invent new execution paths. Dynamic workflows can also use native Python control flow instead of static graph config. Model-agnostic, though the examples unsurprisingly feature Gemini. Where it sits: same territory as [[LangGraph]] (graph-based agent orchestration) and [[Mastra Workflows]], with Google's weight behind it and a harder line on determinism by default. ## References - Why we built ADK 2.0: https://developers.googleblog.com/why-we-built-adk-20 ## Related - [[Use code for control flow, agents for reasoning]] - [[LangGraph]] - [[Mastra Workflows]] - [[AI Agents]] - [[Distinction between AI Agents and Automation Workflows]]