Skip to content

README.md

Latest commit

 

History

History
81 lines (65 loc) · 3.32 KB

README.md

File metadata and controls

81 lines (65 loc) · 3.32 KB
Raw

Google ADK on Temporal

Run Google ADK (adk-go) agents durably on Temporal using the googleadk contrib integration. The agent's orchestration loop runs inside a Workflow; each model call runs as a Temporal Activity (via googleadk.NewModel), and tools can be ordinary Temporal activities exposed to the agent with googleadk.ActivityAsTool — so model calls and tools are retried, timed-out, and visible in the Temporal UI, and the whole run is replayable.

Agents are built the native ADK way (llmagent.New + runner.New); the only Temporal-specific pieces are googleadk.NewModel(...) as the agent's model, googleadk.NewContext(ctx) passed to Run, and the worker-side googleadk.NewPlugin(...) in worker.Options.Plugins, whose config holds the real Gemini client (so the API key stays worker-side, never crossing into the workflow). The tests register the integration's Activities directly with googleadk.NewActivities(...) instead, since the test environments run no plugins.

Every sample runs against a scripted FakeModel in its *_test.go, so go test ./googleadk/... needs no API key or network.

Samples

  • Basic agent — the root files in this directory (workflow.go, worker/, starter/): a single agent that answers a question, calling a get_weather tool implemented as a Temporal activity via googleadk.ActivityAsTool.
  • multiagent/ — a coordinator root agent that delegates to weather and jokes specialist SubAgents via ADK's in-workflow transfer_to_agent.
  • humanintheloop/ — an agent with a sensitive delete_resource tool whose workflow durably waits on a Temporal signal for a human's approval before the tool runs.
  • chat/ — a long-lived, signal-driven conversation on one ADK session that continues-as-new (exporting/importing the session) to keep history bounded.

Prerequisites (for running against a live model)

Running the basic agent

  1. Start a Temporal server (see prerequisites).

  2. In a second terminal, start the worker (blocks until Ctrl+C):

export GEMINI_API_KEY=...
go run googleadk/worker/main.go
  1. In a third terminal, run the starter:
go run googleadk/starter/main.go

The starter asks "What's the weather in San Francisco?"; the agent calls the get_weather tool and answers. You should see a final log line similar to:

2025/12/22 15:07:25 Agent answer: It's currently sunny and about 72°F in San Francisco.

The exact wording comes from the model and will vary. In the Temporal UI you will see the run's googleadk.InvokeModel and get_weather Activities.

Each scenario subdirectory has its own README with run and test instructions.

Test without a live LLM

workflow_test.go (and each scenario's *_test.go) drives the workflow through a scripted FakeModel from the googleadk contrib package, so it needs no API key or network:

go test ./googleadk/...