I have seen a video or two on this and it seemed possible but the details were thin on the ground so I thought Iโd give it a shot today. The rest of this weekโs experiment was a proper learning curve - it seemed simple, except in practice, not so much.
To start I built a BC Driving Assistant Agent in Google Agent Platform Studio and after a lot of fiddling got it deployed and accessible over A2A.
But, as usual, the interesting pieces came from the bits that did not quite work.
๐ ๐ณ๐ฒ๐ ๐๐ต๐ถ๐ป๐ด๐ ๐ ๐ฟ๐ฎ๐ป ๐ถ๐ป๐๐ผ:
- Copilot Studio had trouble finding the Agent Card, even though I could see the Agent Card directly in a browser, and I could see calls from Copilot going to the same URL. That was odd.
- I then manually configured the Google agent inside Copilot Studio. At that point, I could see the Google agent being called, but Copilot started complaining about an API key. Even though I had no authentication configured. (๐๐ฏ๐ฅ ๐บ๐ฆ๐ด, ๐ฃ๐ฆ๐ง๐ฐ๐ณ๐ฆ ๐ข๐ฏ๐บ๐ฐ๐ฏ๐ฆ ๐ด๐ข๐บ๐ด ๐ช๐ต, ๐ต๐ฉ๐ฆ ๐ถ๐ฏ๐ข๐ถ๐ต๐ฉ๐ฆ๐ฏ๐ต๐ช๐ค๐ข๐ต๐ฆ๐ฅ ๐ต๐ฆ๐ด๐ต ๐ด๐ฆ๐ต๐ถ๐ฑ ๐ฉ๐ข๐ด ๐ฏ๐ฐ๐ธ ๐ฃ๐ฆ๐ฆ๐ฏ ๐ต๐ข๐ฌ๐ฆ๐ฏ ๐ฅ๐ฐ๐ธ๐ฏ.)
- Itโs not clear whether a Google AI Agent Studio can be directly deployed with an A2A bridge in front of it, referenced only via the Agent ID, or whether the Agent has to be built and deployed via ADK.
๐ ๐ณ๐ฒ๐ ๐ผ๐ฏ๐๐ฒ๐ฟ๐๐ฎ๐๐ถ๐ผ๐ป๐ ๐ณ๐ฟ๐ผ๐บ ๐๐ต๐ฒ ๐ฝ๐ฟ๐ผ๐ฐ๐ฒ๐๐:
- One slightly painful detour was Gemini sending me looking for a slider or button to โenable A2Aโ on the agent I had built - which doesnโt look like it exists.
- I really liked being able to work at the command line with gcloud shell and the cloud shell editor - this suited my working style much better. I could get to the underlying code and configuration files, make changes, rerun the service, tail logs, and troubleshoot much more directly.
The pattern is now making sense though - itโs clear that purpose-specific agents, running in the environments where they work best, connected through A2A, is a very interesting direction (even across completely different vendors).
Part 1 got me to a partially working bridge and a much better understanding of the moving parts. Part 2, next week, will be about getting the handoff working cleanly, understanding the authentication model properly, and seeing whether this can become a practical cross-platform agent pattern rather than just a Sunday morning science project.
Definitely more to learn - but that was the point.