Observed Signal · Jul 31, 2026 · Technical Release · Source: DEV Community · Impact: 3/5 · Sentiment: Neutral
Google ADK Master Calls Amazon Bedrock via A2A v1.0
A cross-cloud interoperability benchmark reversed master/worker roles: a Google ADK master on Cloud Run (Gemini 2.5 Flash, us-central1) delegated work to an Amazon Bedrock AgentCore worker (Strands Agents on Nova Micro, us-east-1) over A2A v1.0. The reversal exposed six interoperability defects (five invisible to local tests) across protocol versioning, missing server dependencies, identity federation/IAM trust-policy errors, header forwarding through proxies, and agent-card bind-addresses. On 2026-07-31 (five warm invocations per mode) median latencies were mcp_only 2.96s, a2a_only 18.92s, and verified 16.19s, showing the remote-agent hop costs roughly 6× the local-tool baseline; verified mode runs MCP and A2A concurrently so cost approximates the slower call. The article documents fixes and lessons about testing at deployment seams while noting limited sample size, single region pair, and unresolved policy-scoping questions.
Documents cross-cloud agent interoperability failures and latency/cost characteristics for agent delegation between major cloud LLM runtimes (Google and AWS); relevant to cloud AI infrastructure and identity configuration but not directly industry-shifting for AdTech.
Track Google Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Topology: Google ADK master on Cloud Run (Gemini 2.5 Flash, us-central1) called an Amazon Bedrock AgentCore worker (Strands Agents on Nova Micro, us-east-1) over A2A v1.0.
- Reversal revealed six interoperability defects; five could not be detected by local tests alone.
- Measured medians (2026-07-31, five warm runs per mode): mcp_only 2.96 s, a2a_only 18.92 s, verified 16.19 s.
- Remote-agent delegation increased latency ~6× versus the MCP baseline; verified mode executes MCP and A2A concurrently so cost ≈ max(mcp, a2a).
- Root causes included protocol version mismatch (v0.3 vs v1.0), missing server dependency (sse_starlette/http-server extra), IAM/OIDC trust-policy misconfiguration (including explicit Google OIDC provider breaking native federation), dropped A2A-Version header, and agent cards advertising container-local bind addresses.
Connected Companies & Entities
3 Entities mapped“A Google ADK master on Cloud Run (Gemini 2.5 Flash, `us-central1`) owns the benchmark policy....”
“It calls an Amazon Bedrock AgentCore worker (Strands Agents on Nova Micro, `us-east-1`) over A2A v1.0....”
“It calls an Amazon Bedrock AgentCore worker (Strands Agents on Nova Micro, `us-east-1`) over A2A v1.0....”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Cross-Cloud A2A Benchmark: AWS, Azure, GCP Interoperate
The author deployed and measured all six directed cross-cloud A2A (agent-to-agent) paths between three clouds (Microsoft Foundry on Azure, Google ADK on Cloud Run, and AWS/AgentCore). A2A v1.0 interoperated in every deployed direction; the protocol itself was not the primary source of latency or failure. Latency clustered by remote runtime/model (Cloud Run with gemini-2.5-flash ≈1.7–2.1s; Foundry gpt-5-mini ≈18.8–23.4s; AgentCore Nova Micro ≈18.9–25.1s). Verified mode runs concurrently and therefore costs roughly the slower leg (one A2A call), not the sum of both. Recurring cross-cloud failures were identity/authentication, dependency/version skew, packaging/container layout, header forwarding, and timeout policies. The benchmark shows verification provides independent failure detection rather than improved accuracy on the happy path. Publication date: 2026-07-31.
Amazon Bedrock AgentCore A2A Agent Card Contents
The article inspects the A2A agent card published by a Strands agent running on Amazon Bedrock AgentCore Runtime and compares it field-by-field with the same agent deployed on Google Cloud Run and Azure Container Apps. Measurements (storing exact response bytes) were taken on 2026-08-25. Key findings: Bedrock AgentCore requires a separate IAM action GetAgentCard for discovery; the AgentCore card is larger (2,109 bytes) and publishes the agent's full system prompt and model id in the skill description/tags; AgentCore exposes both legacy 0.x fields (url, preferredTransport) and 1.0-style supportedInterfaces (a hybrid shape) and therefore is more compatible with older clients; several optional security/auth fields are missing across all three clouds. The article highlights operational and security implications for agent discovery and client behavior.
Foundry Master, Bedrock AgentCore Smoke Test Passed
A cross-cloud smoke test demonstrated Microsoft Foundry acting as the master coordinator and Amazon Bedrock's AgentCore acting as a remote A2A specialist for a currency conversion workflow. On July 30, 2026, Foundry executed three benchmark modes (MCP-only, A2A-only, and verified), producing identical converted amounts and a deterministic verification that reported a relative_difference of 0. The test required solving cross-cloud authentication (using an OAuth client-credentials flow with AWS Cognito and a custom JWT authorizer), SDK dependency separation, timeout increases for cold A2A calls, and Azure Container Registry authentication for image pulls. Code and sanitized evidence are published in a GitHub repository and the project includes 66 deterministic tests; the result is proof of cross-cloud orchestration coverage, not a full performance benchmark.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.