Last week I got close, but authentication stopped the experiment from working properly. This week I went back in with a slightly different approach.
Quick reminder - the architecture was:
๐๐ผ๐ฝ๐ถ๐น๐ผ๐ ๐ฆ๐๐๐ฑ๐ถ๐ผ ๐๐ด๐ฒ๐ป๐ ๐๐ผ
- Google A2A connection
- Google agent exposed through an A2A bridge
- Cloud Run deployment
- Gemini-based Driver Information Assistant Agent using Google tools.
The Google agent was a simple โ๐๐ณ๐ช๐ท๐ฆ๐ณ ๐๐ด๐ด๐ช๐ด๐ต๐ข๐ฏ๐ต ๐๐จ๐ฆ๐ฏ๐ตโ for British Columbia driving questions, using ICBC (Insurance Corporation of British Columbia) as the source context.
The major change I tried this week was to use Gemini CLI in Google Cloud Shell to help build and deploy the agent - it was simply awesome.
It handled the Cloud Run deployment, found an error, rebuilt, and helped me get to a visible agent card at the expected .๐ธ๐ฆ๐ญ๐ญ-๐ฌ๐ฏ๐ฐ๐ธ๐ฏ/๐ข๐จ๐ฆ๐ฏ๐ต-๐ค๐ข๐ณ๐ฅ.๐ซ๐ด๐ฐ๐ฏ endpoint. The build and deployment logs were so easier to follow than a vague โ๐ด๐ฐ๐ฎ๐ฆ๐ต๐ฉ๐ช๐ฏ๐จ ๐ช๐ด ๐ฉ๐ข๐ฑ๐ฑ๐ฆ๐ฏ๐ช๐ฏ๐จโ status message.
Once the agent card was visible, Microsoft Copilot Studio was able to discover the Google agent and add it as an A2A-connected agent (the first win over last week).
The second issue was the API authentication error. The root cause appears to have been the deployment approach. Using the Gemini API meant the agent expected an API key. Switching to Vertex AI mode (GOOGLE_GENAI_USE_VERTEXAI=TRUE ), and a few other small changes, allowed the deployment to use service account authentication instead. After redeploying with Vertex AI, the Copilot Studio Agent successfully called the Google agent.
A few practical observations:
- Copilot Studio did a better job this time discovering the agent card.
- The raw output in the Copilot test window was genuinely useful for troubleshooting.
- The fallback to the Copilot Studio agentโs own knowledge when A2A failed is worth being aware of, and can be dealt with via configuration.
- The authentication model needs to be designed deliberately before this is used for anything serious.
- The deployment instructions need to be precise, especially around Gemini API versus Vertex AI.
The broad takeaway for me is that A2A is promising and works and for a simple PoC not that difficult at all to get working. ๐๐ฟ๐ผ๐บ ๐ฎ ๐ฆ๐ผ๐น๐๐๐ถ๐ผ๐ป ๐๐ฟ๐ฐ๐ต๐ถ๐๐ฒ๐ฐ๐๐๐ฟ๐ฒ ๐ฝ๐ฒ๐ฟ๐๐ฝ๐ฒ๐ฐ๐๐ถ๐๐ฒ A2A makes cross-platform agent orchestration plausible, but the real architecture work is not the agent conversation itself. It is discovery, identity, failure handling, observability, data boundaries, deployment repeatability, and governance.
While still a weekend experiment, but this one felt like progress and an architecture pattern worthy of further assessment.