Backends & execution

What .build() actually does, what runs the resulting graph, and how to swap the runtime without rewriting the agents.

        flowchart LR
    F[Fluent builder<br/>Agent, Pipeline,<br/>FanOut, Loop] --> IR[IR tree<br/>inspectable,<br/>serialisable]
    IR --> N[Native ADK<br/>LlmAgent, Runner]
    IR --> T[Temporal<br/>durable workflow]
    IR --> X[Custom backend<br/>asyncio / DBOS /<br/>Prefect]

    classDef f fill:#e3f2fd,stroke:#1565c0,color:#0d47a1
    classDef b fill:#fff3e0,stroke:#e65100,color:#bf360c
    class F,IR f
    class N,T,X b
    

Chapters

Chapter	Read when
IR & backends	You want to know what .build() produces under the hood, or you’re implementing a custom backend.
Execution backends	Your agents need durable execution, horizontal scaling, or a scheduler outside the Python process.
Temporal guide	You’ve picked Temporal and want the step-by-step wiring: workflow definitions, activities, signalling.

Tip

You almost always want the default The default ADK backend — local, in-process — handles 95% of production workloads including streaming, sessions, and multi-agent topologies. Reach for Temporal or DBOS only when you need durable execution across process restarts or horizontal fan-out.