Google ADK 2.0 adds graph workflows
Agent Development Kit 2.0 brings a directed-graph runtime, programmatic routing instead of LLM orchestration, and strict state boundaries; Go joins Python.
Google published its reasoning behind Agent Development Kit 2.0 on July 1, a day after shipping ADK Go 2.0. The release replaces LLM-driven orchestration with a structured workflow runtime: agents, tools, and conditional routing become nodes in a directed graph, connected by edges that pass only the data the next node needs.
Three design choices define the version. Routing decisions are programmatic rather than delegated to a model, so a workflow takes the same path for the same input. State is strictly bounded per node, which prevents the context bloat that creeps into long agent runs. And Google argues deterministic execution paths are resistant to prompt injection, since a poisoned input cannot reroute a graph whose edges are fixed in code.
ADK 2.0 is available now in Python and Go, joining earlier Java, TypeScript, and Kotlin support, with a task-collaboration layer for multi-agent coordination and human-in-the-loop steps.
The direction rhymes with what the rest of the ecosystem shipped this quarter: Vercel’s AI SDK 7 also pushed structure into agent code, and enterprise MCP auth reaching stable is the same instinct applied to tool access. If you have been burned by an agent that improvised its way through a workflow, the graph model is the correction, and the freelancer briefing guide explains the discipline that still applies inside each node.