Source: Why we built ADK 2.0 (Google Developers Blog)

Google's rationale for ADK 2.0: move orchestration out of the LLM into deterministic graph and dynamic workflows.

Contents

Summary

Argues that production agents loop, skip business logic or fail opaquely when the model handles routing, scheduling and error handling. ADK 2.0 adds a workflow runtime: deterministic graphs, code-driven dynamic workflows, and collaborative agents that each see only the context they need. An illustrative refund example compares a vanilla agent with a workflow.

Key claims

  • Model-focused fixes (guardrails, skills, prompting) only go so far for orchestration reliability.
  • Graph workflows route with code, not the LLM, which limits context bloat and narrows prompt-injection impact (no edge exists for unauthorised actions).
  • Dynamic workflows use native Python control flow and asyncio for retries and branching.
  • In the post’s illustrative benchmark (gemini-3.5-flash, mock APIs): 5,152 vs 2,265 tokens and 7.2 s vs 5.7 s for agent vs workflow.

Feeds

Suggest changes →