Google released ADK Go 2.0 on June 30, 2026, adding a graph-based workflow engine for building multi-agent applications. The update is not a QA tool by itself, but it matters for teams testing AI agents because it moves more agent behavior into explicit routes, nodes, retries, pauses, and resumable execution.

In the official Google Developers Blog announcement, Google says ADK Go 2.0 adds graph workflows, built-in human-in-the-loop support, dynamic orchestration in Go, LLM agent modes, and a unified node runtime. The ADK 2.0 documentation also states that Go v2.0.0 became generally available on June 30, 2026, while Python v2.0.0 reached GA earlier on May 19, 2026.

What changed

The important shift is from a single autonomous prompt loop toward workflow graphs. Instead of asking an LLM to decide every next step, developers can define nodes for functions, tools, agents, joins, dynamic branches, and parallel workers. Edges and routes decide how work moves through the graph.

Google’s ADK Go 2.0 post highlights several production-focused features: workflows can pause for a human response, resume later, use retry policies with exponential backoff and jitter, cap concurrency, isolate parallel branches, and expose node or agent execution through more consistent telemetry.

Why this matters for QA engineers

QA teams are increasingly asked to test agentic systems, not just web pages or APIs. A graph-based agent workflow gives testers more concrete behavior to verify: which route fired, which node ran, whether a human approval pause happened, what state was persisted, and whether retries stopped at the expected limit.

  • Deterministic routes: QA can create tests for known workflow paths instead of only judging final chat output.
  • Human-in-the-loop pauses: approval, correction, and resume flows become first-class test cases.
  • Retry behavior: testers can validate backoff, timeout, and failure handling when tools or APIs are unstable.
  • State boundaries: teams can check that one branch does not leak irrelevant context into another agent’s prompt.
  • Prompt-injection controls: fixed workflow edges can reduce the chance that a malicious input triggers an unauthorized action, but QA still needs negative tests.

QA checks to add for ADK-style agents

If your engineering team starts using ADK 2.0 or a similar graph-first agent framework, ask for test evidence at the workflow level. Final answers are not enough.

  1. Map each business rule to the expected workflow route or node.
  2. Test the happy path, every conditional branch, and at least one invalid route.
  3. Force a tool failure and confirm retries, timeout behavior, and user-visible errors.
  4. Pause the run for human approval, resume it, and verify state is not lost.
  5. Inject irrelevant or malicious text into user input and confirm restricted actions cannot run without the required edge, node, or approval.
  6. Review logs or traces to confirm what actually executed, not only what the agent summarized.

Bottom line

Google ADK Go 2.0 is another signal that AI agents are becoming structured software systems. For QA engineers, that is useful news: graph workflows create clearer places to test routing, state, retries, approvals, and security boundaries. The risk is assuming the graph makes the agent correct automatically. It does not. It gives QA teams better surfaces to verify.

Sources