English
What problem does this solve?
An agent that works perfectly on the developer's own machine tells you almost nothing about whether it will work anywhere else. Different Python versions, missing system libraries, or an environment variable that only exists locally are exactly the kind of gap that turns "it works for me" into a production incident. This walkthrough continues studying the open-source theailanguage/a2a_samples reference implementation.
How the mechanism works
This part first extends the agent itself: a Vision Agent that accepts a question plus an image path, and has Gemini describe or analyze what's in the picture — proving the same A2A task pattern works for multimodal input, not just text.
Then the whole application — operating system layer, dependencies, and source code — gets packaged into a single Docker image, so it behaves identically wherever it runs. The build is tested locally first: build the image, run the container, and send it a real request over curl before trusting it with anything further. Only after that local check passes does the image get pushed to Google's container registry and deployed to Cloud Run with a single command, returning a real public HTTPS URL.
Trade-offs and alternatives
The benefit is a deployment that is genuinely portable and reproducible — the same image that passed the local curl test is the exact image running in the cloud, not a "close enough" recreation. The trade-off is the extra step of learning Docker and container registries, which is real overhead for a team that has never containerized anything before.
The alternative — deploying source code directly to a platform-specific runtime without a container — can be faster to set up initially, but it reintroduces the "works on my machine" risk the moment that platform's environment drifts even slightly from the developer's.
Conclusion
Local success and production readiness are two different claims. Testing the exact container that will ship — not just the code — is what closes that gap, and it's the same discipline that applies whether the agent is answering text questions or now, images too.