Running in production¶

The package builds an ADK App. Everything around it (sessions, servers, users, scaling) works the way ADK does. This page covers the points where the package's features meet that infrastructure.

Sessions¶

Use a persistent ADK session service, such as DatabaseSessionService for Postgres (install SQLAlchemy and an async driver like asyncpg in your app). The task ledger, prebuilt and foreground subagent history, and pending questions to the user are all stored in the session.

Run one turn per session at a time. ADK's database session services reject writes to a session that changed after it was loaded, so queue new messages for a session until its current turn finishes.

Users and sandboxes¶

A sandbox belongs to the App, and every session of that App shares it. For per-user isolation, create one App and Runner per user with that user's environment, cache them, and call runner.close() when the user is idle or done, which removes the sandbox. See Sandbox workspaces.

Background subagents¶

Background tasks are tracked in the memory of the process that started them:

• route a session's requests to the same process (sticky sessions) while it has background tasks running;
• expect running tasks to be lost on restart or deploy;
• forward each task's events and start the "task finished" turn from that process. See the example.

Asking the user¶

elicitation=True makes the App resumable. Keep the invocation_id from the ElicitationRequest with your pending UI state, so the answer can resume the interrupted run, even after a page reload.

Long sessions¶

• Set compaction with a token_threshold well below your model's context window.
• Consider context_cache when instructions, skills, and tools are large and stable.
• event.usage_metadata reports token counts per model call if you want to track cost or context use.

Versions¶

The package depends on google-adk>=2.7.1,<3 and uses some ADK features still marked experimental: EnvironmentToolset, the Daytona and E2B environments, and resumable Apps. Pin the versions you've tested in your app and upgrade deliberately.