Amazon Bedrock Agents Orchestrating Google ADK over A2A

An Amazon Bedrock master agent, built with Strands Agents and hosted on AWS, delegates to a Google ADK worker on GCP Cloud Run over the A2A v1.0 protocol. The master cross-checks the worker against an MCP exchange-rate tool and measures latency, reliability, and failure behavior of cross-cloud verification. The benchmark also compares performance with a previous run using Microsoft Foundry in Azure.

This article explains how to build and test a cross-cloud currency agent. An Amazon Bedrock master agent , built with Strands Agents and hosted on Amazon Bedrock AgentCore Runtime in AWS us-east-1 , delegates to a Google ADK worker on GCP Cloud Run in us-central1 over A2A v1.0 . The master cross-checks the worker against an MCP exchange-rate tool and measures the latency, reliability, and failure behavior of cross-cloud verification. Most Agent-to-Agent A2A protocol demos stop at "look, the HTTP 200 OK request succeeded." That is a smoke test, not an interoperability benchmark. This project goes further: the Bedrock master owns the user interaction and benchmark policy. It discovers and calls a Google ADK worker running on GCP Cloud Run, then compares the worker result with a local MCP stdio exchange-rate tool backed by live Frankfurter daily reference rates. We also compare the performance, developer experience, and wire compatibility with a previous benchmark run using Microsoft Foundry in Azure gpt-5-mini . Together, the runs cover components hosted across AWS, Azure, and GCP. The benchmark addresses four questions: This builds directly on the currency agent from the previous articles in this series: That agent — built with Google ADK, Gemini 2.5 Flash, and a FastMCP exchange-rate server backed by the free Frankfurter API https://www.frankfurter.dev/ — serves as the remote worker and independent verifier in this project. The new repository adds the AgentCore coordinator and benchmark suite: CLI / Boto3 Test Runner AWS SigV4 Auth | Bedrock AgentCore Runtime hosted master AWS, us-east-1, Amazon Nova Micro Strands Agents orchestration | +-- MCP stdio -- Frankfurter rates in-container stdio process | +-- A2A v1.0 -- Cloud Run GCP, us-central1 | Google ADK worker gemini-2.5-flash | MCP HTTP -- Frankfurter rates The Bedrock master answers every conversion request through three distinct evaluation modes: | Mode | What happens | Why it exists | |---|---|---| mcp only | Bedrock master calls the local MCP rate tool | Baseline single-agent performance | a2a only | Bedrock master delegates to the GCP ADK worker over A2A v1.0 | Measure remote-agent behavior and network latency | verified | MCP result independently checked against the remote ADK agent over A2A | Measure the accuracy-versus-overhead tradeoff | Both sides read the same Frankfurter daily reference rates on purpose: when the two clouds disagree, that measures protocol, model, and orchestration behavior, not data-source skew. Currency conversion is a poor job for an LLM and a good job for Python's Decimal . The domain layer is framework-independent and uses Pydantic models. Numeric agreement is evaluated in code using relative difference; no LLM is asked, "Do these numbers look close to you?" difference = abs primary.converted amount - verifier.converted amount relative = difference / abs primary.converted amount agreed = relative <= tolerance default 0.005 0.5% The failure policy is explicit rather than emergent: unverified . "verification unavailable" warning. validation , provider , authentication , transport , timeout , protocol ; never fabricate a rate.Because "which layer broke" is a core research question, every adapter exception is normalized into exactly one typed failure at the boundary. The first attempt to connect the AgentCore coordinator to the Google ADK currency agent died immediately on invocation: a2a.utils.errors.MethodNotFoundError: Method not found Observed root cause: a protocol-version mismatch between A2A v0.3.0 and v1.0, with no automatic fallback negotiation in the tested client. a2a-sdk =1.0 calls the A2A v1.0 JSON-RPC method SendMessage . a2a-sdk 0.3.x only expose the v0.3.0 method message/send . protocolVersion: 0.3.0 — but attempted the v1.0 method anyway.The initial ecosystem package pins were also mutually exclusive: | Package | a2a-sdk Requirement | Status | |---|---|---| strands-agents 1.50.2 | =1.0.0,<2 | Compatible | google-adk 2.1.0 – 2.4.0 | =0.3.4,<0.4 | Incompatible | google-adk 2.5.0 | =0.3.4,<2 | Compatible ✅ | a2ui-agent-sdk through 0.4.0 | <0.4.0 | Incompatible ❌ | google-adk 2.5.0 updated its dependencies to support a2a-sdk 1.x . However, A2UI extensions currently pin the older v0.3.0 protocol. For this benchmark, A2UI was omitted so both AWS and GCP sides could operate on A2A v1.0 a2a-sdk 1.1.2 . Deploying the master to Amazon Bedrock AgentCore Runtime involved navigating several fast-moving SDK and platform details observed during our build on 2026-07-28: In the account used for this build, Anthropic models such as Claude 3.5 Sonnet required a one-time use-case submission PutUseCaseForModelAccess and an AWS Marketplace subscription agreement. To keep the setup automated, we configured the coordinator to use Amazon Nova Micro us.amazon.nova-micro-v1:0 . Nova Micro required no approval form in our test account, supported native tool calling in the tested workflow, and produced subsecond model responses. In our deployment, using the bare model ID amazon.nova-micro-v1:0 returned an HTTP 400 ValidationException requiring on-demand throughput configuration. Passing the regional inference profile ID us.amazon.nova-micro-v1:0 resolved the error. The older Python pip -based starter toolkit agentcore configure / agentcore launch was deprecated in June 2026. Deployment now uses the official @aws/agentcore npm CLI Node 20+, CDK-based . app/CurrencyCoordinator/main.py python from bedrock agentcore.runtime import BedrockAgentCoreApp from strands import Agent, tool from coordinator.hosted tool import run currency benchmark from model.load import load model app = BedrockAgentCoreApp tools = tool run currency benchmark The full source defines a bounded, session-scoped agent factory here. @app.entrypoint async def invoke payload, context : session id = getattr context, "session id", "default-session" agent = get or create agent session id prompt = payload.get "prompt", payload.get "messages", "" result = await agent.invoke async prompt return {"result": str result } if name == " main ": app.run The hosted runtime also fails closed when its GCP worker is missing: { "name": "CURRENCY REQUIRE GCP ADK", "value": "1" } With that setting, a2a only and verified return gcp adk not configured if CURRENCY A2A ENDPOINT is absent. A deployment can no longer appear to exercise A2A while silently using a local fixture. The Bedrock model configuration also sets BEDROCK MAX TOKENS=1024 explicitly to bound output and quota usage. The remote verifier container colocates two processes: the FastMCP Frankfurter server on localhost and the A2A app listening on $PORT . Gemini API keys are retrieved securely from GCP Secret Manager: gcloud secrets create gemini-api-key --data-file="$HOME/gemini.key" gcloud run deploy currency-adk-a2a \ --source adk agent --region us-central1 \ --allow-unauthenticated --min-instances=0 --max-instances=2 \ --set-secrets "GOOGLE API KEY=gemini-api-key:latest" \ --set-env-vars "MCP SERVER URL=http://127.0.0.1:8081/mcp,GENAI MODEL=gemini-2.5-flash" Setting --min-instances=0 allows Cloud Run to scale to zero when idle. The coordinator's timeout is set to 60 seconds to accommodate initial container cold starts. The repository includes a complete local test suite that runs deterministically without credentials or cloud infrastructure: 1. Clone & install dependencies git clone https://github.com/xbill9/bedrock-adk-a2a-currency cd bedrock-adk-a2a-currency pip3 install --user -e ". dev " 2. Run unit and integration tests deterministic fixtures pytest 3. Test local CLI modes currency-benchmark 100 USD CAD EUR --mode mcp only currency-benchmark 100 USD CAD EUR --mode verified --transport mcp-stdio 4. Execute full evaluation matrix currency-evaluate --output /tmp/currency-results.jsonl --summary /tmp/currency-summary.json To deploy and test the hosted Bedrock master: ./infra/sync app.sh agentcore deploy -y agentcore invoke "Convert 100 USD to EUR and CHF in verified mode." On 2026-07-29, I deployed the updated master to AgentCore Runtime in us-east-1 and invoked all three modes through the hosted InvokeAgentRuntime API: | Hosted mode | Observed result | |---|---| mcp only | HTTP 200; live mcp-stdio:frankfurter-live quote | a2a only | HTTP 200; live gcp-adk-a2a-worker quote | verified | HTTP 200; MCP and GCP ADK agreed exactly for EUR and CHF | The verified request converted 100 USD to EUR and CHF. The deterministic comparison recorded relative difference: "0" and agreed: true for both currencies, with no failures or warnings. The benchmark tool completed in approximately 3.08 seconds. This was an end-to-end smoke test, not a full hosted latency distribution. It exercised the complete path: AWS SigV4 invocation → AgentCore Runtime → Nova Micro tool selection → MCP stdio / Frankfurter → A2A v1.0 → GCP Cloud Run → Google ADK / Gemini → deterministic Decimal comparison The smoke test also found a real orchestration bug. On the first request, Nova Micro read “Convert 100 USD to EUR” but claimed the target currency was missing and asked the user to confirm it. The master prompt now includes an explicit natural-language parsing rule and forbids confirmation requests for information already present. After redeployment, the same request called the benchmark tool directly. A regression test preserves that behavior. We executed the 38-case evaluation matrix across all three modes: 114 records per run. The 2026-07-28 warm run exercised the framework-independent coordinator locally against the live GCP Cloud Run ADK endpoint; it did not measure the AgentCore hosting layer. The 2026-07-27 run is the retained Azure-era baseline. Keeping those labels explicit avoids attributing local harness latency to AgentCore. | Observed run | Evaluation mode | Success rate | Median latency | p95 latency | Agreement rate | |---|---|---|---|---|---| | 2026-07-28 warm local harness → GCP | mcp only | 100% | 286 ms | 540 ms | N/A | | 2026-07-28 warm local harness → GCP | a2a only | 100% | 2.09 s | 6.10 s | N/A | | 2026-07-28 warm local harness → GCP | verified | 100% | 1.87 s | 4.33 s | 96.77% | | 2026-07-27 Azure-era baseline → GCP | mcp only | 100% | 297 ms | 1.09 s | N/A | | 2026-07-27 Azure-era baseline → GCP | a2a only | 100% | 1.69 s | 4.82 s | N/A | | 2026-07-27 Azure-era baseline → GCP | verified | 100% | 1.71 s | 4.15 s | 96.77% | message/send and v1.0 SendMessage are wire-incompatible. If you see MethodNotFoundError , inspect the a2a-sdk version on both client and server before debugging prompt logic. us.amazon.nova-micro-v1:0 avoided the on-demand throughput error returned for the bare model ID. Decimal arithmetic prevents LLM calculation errors from affecting conversion and agreement checks. The checks therefore measure differences in returned results, not the model's arithmetic ability. mcp only path is useful as a baseline, while cross-cloud A2A verification adds an independent result for failover and anomaly detection. Whether the overhead is justified depends on the workload.The complete benchmark codebase, deployment scripts, test suite, and raw evaluation datasets are available on GitHub: If you are building multi-cloud agent systems with Amazon Bedrock AgentCore, Google ADK, or Microsoft Agent Framework, feedback and benchmark contributions are welcome.