The problem ADK 2.0 tries to solve
Until now, building a reliable agent in production has meant spending most of your time containing the model's creativity: writing retry boilerplate, managing persistent state, deciding when to let the LLM reason freely and when to force it into a predictable sequence. ADK 2.0 — announced at the I/O 2026 developer keynote — tries to make those choices part of the framework rather than something teams reinvent every time.
Workflow Runtime
The centerpiece is the Workflow Runtime: a graph-based execution engine that composes deterministic flows for agentic apps. It supports routing, fan-out/fan-in, loops, retries, state management, dynamic nodes, human-in-the-loop and nested workflows. The developer chooses, case by case, where to set the slider: model-led reasoning where flexibility matters, fixed sequences where guarantees do.
Multi-Agent Collaboration — previously known as the Task-based Agent Collaboration API — formalizes teams of agents: a coordinator delegates to subagents through explicit modes (chat for full user interaction, task with clarifications, single-turn for parallel execution without a user in the loop).
Agents CLI and the ecosystem
Google also released the Agents CLI: it packages ADK skills — eval, deploy, observability, publishing — so any AI coding agent becomes an expert at building agentic apps. It integrates with AG-UI and A2UI, the new communication protocols introduced for agentic experiences on Android.
Why it matters
For anyone weighing whether to roll their own orchestration or rely on a framework, ADK 2.0 raises the floor of what Google's open-source baseline gives you without writing infrastructure code. For enterprise teams — as Virtualization Review notes — it pairs with the Gemini Enterprise Agent Platform and Managed Agents, completing the "reasoning + execution + governance" stack.
Available on GitHub and in the Google Cloud documentation.