Google Cloud details graph workflows in ADK: fan-out, routers, and human-in-the-loop
rseroter · x · 2026-09-30
Google Cloud's Annie Wang and Shangjie Chen published a hands-on guide to graph workflows in the Agent Development Kit (ADK).
- Graph engineering means breaking a task into nodes, connecting them with edges, and deciding whether code, models, or humans control the next step
- Using a refund workflow example, the post covers fan-out/fan-in parallelism, deterministic and agent-based routers, human-in-the-loop pauses, and parallel workers processing lists of cases
- They start from a single all-in-one agent, then show how splitting the prompt's implicit steps (lookups, decision, reply) into separate nodes gives per-step control
- Guidance on when to declare a static graph versus letting Python dynamically schedule work as results arrive
Takeaway: starting with one agent is fine, but splitting into single-purpose agents with graph orchestration gives more control as complexity grows.
More from coding & agent
- Hugging Chat's ML Intern adds MCP OAuth, virtual filesystem and downloadable traces — Gradio · 2026-09-30
- CSS tip: flip overscroll-behavior default with any-pointer to keep pull-to-refresh on touch devices — jh3yy · 2026-09-30
- Scenario open-sources full recipe: 2 drawings yield 4-direction isometric walking spritesheets — kastnerkyle · 2026-09-30
- Hillel Wayne: TLA+ is great, but 'formal methods will save AI' hype misses what it can't even express — leland_mcinnes · 2026-09-30
- Decision model Jev makes web agents 4.3x faster and 13x cheaper across 240 runs — EdenEmarco177 · 2026-09-30
- Coinbase CEO unveils financial accounts built for AI agents and superintelligence — J0se · 2026-09-30