DEV Community

AWS Community Builders profile image xbill
xbill for AWS Community Builders

Posted on

Cross Cloud A2A Agent Benchmarking

#a2aprotocol #gcp #aws #mcps

Building a Benchmarking Agent with A2A and MCP

This tutorial aims to build and test benchmarking Agents using the A2A protocol across several mainstream Cloud providers. A Master Orchestrator Agent is exposed via MCP to allow Antigravity CLI to be used as a MCP client to co-ordinate the benchmarks.

Deja Vu — What is Old is New!

This paper is a re-visiting of the original benchmark series with Gemini CLI over Node, GO, and Python:

Cross Language A2A Agent Benchmarking with Gemini 3 and Gemini CLI

In this updated version, the Antigravity CLI is used to push Rust Agents cross-cloud and co-ordinate Mersenne Prime Calculations.

Why would I need Multi-Cloud Support? And Rust? Can’t I just use Python?

Most mature Agent development tools and libraries are Python based. Python allows for rapid prototyping and evaluation of approaches. Python is also an interpreted language- which has trade-offs in memory safety, and performance. Other languages like GO and Rust offer high performance and memory safe operations. With a language neutral communication protocol — the actual Agent implementation of each Agent can be coded in the most appropriate language.

What is this Approach actually Benchmarking?

The high level goal was to measure the actual time spent running an algorithm in the native language code inside the A2A agent. Each language had a slightly different implementation due to the language syntax. After running the algorithm- each Agent was instructed to calculate and return the elapsed time for cross cloud comparison.

What is the A2A protocol?

The Agent2Agent (A2A) protocol, an open communication standard for AI agents, was initially introduced by Google in April 2025. It is specifically engineered to facilitate seamless interoperability within multi-agent systems, enabling AI agents developed by diverse providers or built upon disparate AI agent frameworks to communicate and collaborate effectively.

A good overview of the A2A protocol can be found here:

A2A Protocol

Language Support For the A2A Protocol

The official ADK for Python, GO, and Java provide built-in support for working with the A2A protocol. For other programming languages like JS, Rust, and .NET — 3rd party libraries are available to add support for the protocol.

The main source for A2A Language support is the GitHub repo:

GitHub - a2aproject/A2A: An open protocol enabling communication and interoperability between opaque agentic applications.

Tool Chain Setup for A2A

This article targets a local, GCP, Azure, and AWS environment. The build tools for each language and cloud environment need to be in place. For building with Rust, the Rust tool chain is required.

For GCP- this article provides some directions:

Deploying a Rust A2A Agent to Google Cloud Run

For Azure — this article provides some directions:

Deploying a Rust MCP Server to Azure ACA

For AWS — this article provides some directions:

Deploying a Rust A2A Agent to AWS Lightsail

Antigravity CLI

Antigravity CLI is the follow-on successor to Gemini CLI- the terminal driven, agent assisted coding tool.

Full details on installing Antigravity CLI are here:

Getting Started with Antigravity CLI

Testing the Antigravity CLI Environment

Once you have all the tools in place- you can test the startup of Antigravity CLI.

You will need to authenticate with a Google Cloud Project or your Google Account:

agy
Enter fullscreen mode Exit fullscreen mode

This will start the interface:

Checking the Developer Environment

Verify that all the prerequisite packages and compilers are installed — and clone the sample Github repo:

git clone https://github.com/xbill9/a2a-multicloud
cd a2a-multicloud
Enter fullscreen mode Exit fullscreen mode

Once you have your Google Cloud Project and preferred authentication method — run the init.sh script to validate the setup:

xbill@penguin:~/a2a-multicloud$ source init.sh
WARNING: Your active project does not match the quota project in your local Application Default Credentials file. This might result in unexpected quota issues.

To update your Application Default Credentials quota project, use the `gcloud auth application-default set-quota-project` command.
[environment: Development]
Updated property [core/project].
ADC is valid.
Environment setup
GOOGLE_GENAI_USE_VERTEXAI=1
GOOGLE_CLOUD_PROJECT=comglitn
GOOGLE_CLOUD_LOCATION=us-central1
IMAGEN_MODEL="imagen-3.0-fast-generate-001"
GENAI_MODEL="gemini-2.5-flash"
Cloud Login
  Credentialed Accounts
ACTIVE ACCOUNT
* xbill@glitnir.com

To set the active account, run:
    $ gcloud config set account `ACCOUNT`
Enter fullscreen mode Exit fullscreen mode

The set_env.sh script is provided to set common ADK environment variables:

xbill@penguin:~/a2a-multicloud$ source set_env.sh
Current Environment
export GOOGLE_GENAI_USE_VERTEXAI=1
export GOOGLE_CLOUD_PROJECT=comglitn
export GOOGLE_CLOUD_LOCATION=us-central1
export IMAGEN_MODEL="imagen-3.0-fast-generate-001"
export GENAI_MODEL="gemini-2.5-flash"
export GCP_AGENT_URL="https://bench-rust-289270257791.us-central1.run.app"
export AWS_AGENT_URL="https://a2a-lightsail-rust-aws.6wpv8vensby5c.us-east-1.cs.amazonlightsail.com"
export AZURE_AGENT_URL="https://a2a-app-penguin.icyplant-a768d75c.westus2.azurecontainerapps.io"
Cloud Login
  Credentialed Accounts
ACTIVE ACCOUNT
* xbill@glitnir.com

To set the active account, run:
    $ gcloud config set account `ACCOUNT`

ADK Version
adk, version 2.1.0
Enter fullscreen mode Exit fullscreen mode

Debugging API Permission Errors

If your application default credentials expires or your Google Cloud Authentication expires you will get an error. The workaround is to re-authenticate:

gcloud auth login
gcloud auth application-default login
Enter fullscreen mode Exit fullscreen mode

Another common error is that the environment variables are not set correctly. Go the the root directory and re-run the set_env.sh to set the variables:

cd ~/a2a-multicloud
source set_env.sh
Enter fullscreen mode Exit fullscreen mode

A2A Debugging Tools -A2A Inspector

The A2A Inspector is a standalone tool that provides low level visibility into the A2A protocol. The GitHub is available here:

GitHub - a2aproject/a2a-inspector: Validation Tools for A2A Agents

A summary of the features of the A2A inspector can be found here:

A2A Protocol Documentation

To install the A2A Inspector:

cd ~
git clone https://github.com/a2aproject/a2a-inspector
Enter fullscreen mode Exit fullscreen mode

Then follow the build instructions — you need uv , and a recent version of node :

a2a-inspector/README.md at main · a2aproject/a2a-inspector

Staring the A2A Inspector

Once the A2A inspector has been installed- you can validate the installation by using this URL:

http://127.0.0.1:5001/

Mersenne Primes

In mathematics, a Mersenne prime is a prime number that is one less than a power of two. As of 2025, 52 Mersenne primes are known. An interesting thing about Mersenne primes is that they are the easiest natural numbers to prove to be primes, so they make up the largest category on the list of known prime numbers. This example was chosen as a good problem as the computation is CPU bound and gets exponentially longer for each higher number.

A2A Cross Cloud Agents

The a2a-multicloud Repo has sample scripts for running the agents across cloud environments. Each agent implements a Mersenne Prime Number generator with only basic optimization. The root directory of the a2a-multicloud directory contains several common agent development languages. These include:

  • benchmark-rust (Rust) Local A2A Server
  • benchmark-rust-gcp (Rust) Cloud run A2A Server
  • benchmark-rust-azure (Rust) Azure Hosted A2A Server
  • benchmark-rust-aws (Rust) AWS Hosted A2A server
  • rust-master (Rust) Local A2A Client / MCP Server

A2A Mersenne Prime Number Agent (Rust Local)

This agent provides a basic Agent that generates Mersenne prime numbers.

The A2A inspector can be used to validate the Agent:

The agent can be reviewed:

xbill@penguin:~/a2a-multicloud/benchmark-rust$ make start
Building the Rust project...
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 0.05s
Starting local agent on port 8104 in the background...
Local agent started successfully (PID: 87753).

xbill@penguin:~/a2a-multicloud/benchmark-rust$ make status
Local agent is running (PID: 87753).
Listening on port: 8104
Testing endpoint health...
OK
Enter fullscreen mode Exit fullscreen mode

and the agent card:

xbill@penguin:~/a2a-multicloud/benchmark-rust$ make card
Fetching local benchmark agent card from http://localhost:8104...
{
  "name": "Mersenne Prime Agent Rust",
  "description": "A rust agent that builds a list of the first n Mersenne primes and reports the elapsed time. Configured with model: Not specified.",
  "protocolVersion": "0.3.0",
  "version": "0.1.0",
  "url": "http://0.0.0.0:8104",
  "skills": [
    {
      "id": "find-mersenne-rust",
      "name": "Find Mersenne Primes in rust",
      "description": "Finds the list of the first n Mersenne primes in Rust, or calculates a specific Mersenne prime for a given exponent parameter.",
      "tags": [
        "math",
        "benchmark"
      ]
    },
    {
      "id": "check-status-rust",
      "name": "Check calculation status in Rust",
      "description": "Returns 'ready' if no calculation is active, and 'not ready' if a calculation is live.",
      "tags": [
        "status",
        "health"
      ]
    }
  ],
  "capabilities": {},
  "defaultInputModes": [],
  "defaultOutputModes": []
}
Enter fullscreen mode Exit fullscreen mode

and local A2A call:

xbill@penguin:~/a2a-multicloud/benchmark-rust$ make a2a
Sending A2A status request to http://localhost:8104/...
{
  "id": 1,
  "jsonrpc": "2.0",
  "result": {
    "contextId": "status-context-id",
    "kind": "message",
    "messageId": "971ff86d-7bfc-4046-858a-16e7857ba9d0",
    "parts": [
      {
        "kind": "text",
        "text": "ready"
      }
    ],
    "role": "agent"
  }
}
Enter fullscreen mode Exit fullscreen mode

A2A Mersenne Prime Number Agent (Rust GCP)

This agent provides Rust agent that implements the Mersenne generation algorithm Hosted on GCP.

xbill@penguin:~/a2a-multicloud/benchmark-rust-gcp
make deploy

xbill@penguin:~/a2a-multicloud/benchmark-rust-gcp$ make endpoint
https://bench-rust-fgasxpwzoq-uc.a.run.app
Enter fullscreen mode Exit fullscreen mode

Then get the Agent Card:


xbill@penguin:~/a2a-multicloud/benchmark-rust-gcp$ make status
Checking Google Cloud Run service status for bench-rust...
NAME STATUS URL
bench-rust True https://bench-rust-fgasxpwzoq-uc.a.run.app
Checking deployed agent health at https://bench-rust-fgasxpwzoq-uc.a.run.app/.well-known/agent.json ...
Agent health check: PASSED (HTTP 200)
Enter fullscreen mode Exit fullscreen mode

To check the Agent Card:

xbill@penguin:~/a2a-multicloud/benchmark-rust-gcp$ make card
Fetching agent card from https://bench-rust-fgasxpwzoq-uc.a.run.app/.well-known/agent-card.json ...
{
  "name": "Mersenne Prime Agent Rust",
  "description": "A rust agent that builds a list of the first n Mersenne primes and reports the elapsed time. Configured with model: Not specified.",
  "protocolVersion": "0.3.0",
  "version": "0.1.0",
  "url": "http://0.0.0.0:8080",
  "skills": [
    {
      "id": "find-mersenne-rust",
      "name": "Find Mersenne Primes in rust",
      "description": "Finds the list of the first n Mersenne primes in Rust, or calculates a specific Mersenne prime for a given exponent parameter.",
      "tags": [
        "math",
        "benchmark"
      ]
    },
    {
      "id": "check-status-rust",
      "name": "Check calculation status in Rust",
      "description": "Returns 'ready' if no calculation is active, and 'not ready' if a calculation is live.",
      "tags": [
        "status",
        "health"
      ]
    }
  ],
  "capabilities": {},
  "defaultInputModes": [],
  "defaultOutputModes": []
}
Enter fullscreen mode Exit fullscreen mode

To check that that remote agent can support A2A:

xbill@penguin:~/a2a-multicloud/benchmark-rust-gcp$ make a2a
Sending A2A status request to https://bench-rust-fgasxpwzoq-uc.a.run.app/ ...
{
  "id": 1,
  "jsonrpc": "2.0",
  "result": {
    "contextId": "status-context-id",
    "kind": "message",
    "messageId": "108d0f7c-c2e6-4d1e-8306-843a6dcb999c",
    "parts": [
      {
        "kind": "text",
        "text": "ready"
      }
    ],
    "role": "agent"
  }
}
Enter fullscreen mode Exit fullscreen mode

A2A Mersenne Prime Number Agent (Rust Azure)

This agent provides Rust agent that implements the Mersenne generation algorithm Hosted on Azure:

xbill@penguin:~/a2a-multicloud/benchmark-rust-azure$ make deploy
All requirements met.
Building the Docker image...

xbill@penguin:~/a2a-multicloud/benchmark-rust-azure$ make endpoint
https://a2a-app-penguin.icyplant-a768d75c.westus2.azurecontainerapps.io
Enter fullscreen mode Exit fullscreen mode

Then get the Agent Card:


xbill@penguin:~/a2a-multicloud/benchmark-rust-azure$ make status
mcp-aca-rust-azure is not running locally.
Checking Azure Container App status for a2a-app-penguin...
Name ProvisioningState FQDN
--------------- ------------------- ---------------------------------------------------------------
a2a-app-penguin Succeeded a2a-app-penguin.icyplant-a768d75c.westus2.azurecontainerapps.io
Checking deployed agent health at https://a2a-app-penguin.icyplant-a768d75c.westus2.azurecontainerapps.io/.well-known/agent.json ...
Agent health check: PASSED (HTTP 200)
Enter fullscreen mode Exit fullscreen mode

To check the Agent Card:

xbill@penguin:~/a2a-multicloud/benchmark-rust-azure$ make card
Fetching agent card from https://a2a-app-penguin.icyplant-a768d75c.westus2.azurecontainerapps.io/.well-known/agent-card.json ...
{
  "name": "Mersenne Prime Agent Rust",
  "description": "A rust agent that builds a list of the first n Mersenne primes and reports the elapsed time. Configured with model: Not specified.",
  "protocolVersion": "0.3.0",
  "version": "0.1.0",
  "url": "http://0.0.0.0:8080",
  "skills": [
    {
      "id": "find-mersenne-rust",
      "name": "Find Mersenne Primes in rust",
      "description": "Finds the list of the first n Mersenne primes in Rust, or calculates a specific Mersenne prime for a given exponent parameter.",
      "tags": [
        "math",
        "benchmark"
      ]
    },
    {
      "id": "check-status-rust",
      "name": "Check calculation status in Rust",
      "description": "Returns 'ready' if no calculation is active, and 'not ready' if a calculation is live.",
      "tags": [
        "status",
        "health"
      ]
    }
  ],
  "capabilities": {},
  "defaultInputModes": [],
  "defaultOutputModes": []
}
Enter fullscreen mode Exit fullscreen mode

To check that that remote agent can support A2A:

xbill@penguin:~/a2a-multicloud/benchmark-rust-azure$ make a2a
Sending A2A status request to https://a2a-app-penguin.icyplant-a768d75c.westus2.azurecontainerapps.io/...
{
  "id": 1,
  "jsonrpc": "2.0",
  "result": {
    "contextId": "status-context-id",
    "kind": "message",
    "messageId": "5d6605f6-9926-481f-b807-8598b81b8cc1",
    "parts": [
      {
        "kind": "text",
        "text": "ready"
      }
    ],
    "role": "agent"
  }
}
Enter fullscreen mode Exit fullscreen mode

A2A Mersenne Prime Number Agent (Rust AWS)

This agent provides Rust agent that implements the Mersenne generation algorithm Hosted on AWS:

xbill@penguin:~/a2a-multicloud/benchmark-rust-aws$ make deploy
Building the Docker image...
[+] Building 3.8s (10/13)    

xbill@penguin:~/a2a-multicloud/benchmark-rust-aws$ make endpoint
https://a2a-lightsail-rust-aws.6wpv8vensby5c.us-east-1.cs.amazonlightsail.com/
Enter fullscreen mode Exit fullscreen mode

Then get the Agent Card:


xbill@penguin:~/a2a-multicloud/benchmark-rust-aws$ make status
Checking AWS Lightsail service status for a2a-lightsail-rust-aws...
------------------------------------------------------------------------------------------------------------------------
| GetContainerServices |
+------------+--------+------------+-----------------------------------------------------------------------------------+
| Deployment | Power | State | URL |
+------------+--------+------------+-----------------------------------------------------------------------------------+
| ACTIVE | nano | DEPLOYING | https://a2a-lightsail-rust-aws.6wpv8vensby5c.us-east-1.cs.amazonlightsail.com/ |
+------------+--------+------------+-----------------------------------------------------------------------------------+
Checking endpoint health at https://a2a-lightsail-rust-aws.6wpv8vensby5c.us-east-1.cs.amazonlightsail.com/health...
Health status: HEALTHY (HTTP 200)
Enter fullscreen mode Exit fullscreen mode

To check the Agent Card:

xbill@penguin:~/a2a-multicloud/benchmark-rust-aws$ make card
Fetching agent card from https://a2a-lightsail-rust-aws.6wpv8vensby5c.us-east-1.cs.amazonlightsail.com/.well-known/agent-card.json ...
{
  "name": "Mersenne Prime Agent Rust",
  "description": "A rust agent that builds a list of the first n Mersenne primes and reports the elapsed time. Configured with model: Not specified.",
  "protocolVersion": "0.3.0",
  "version": "0.1.0",
  "url": "http://0.0.0.0:8080",
  "skills": [
    {
      "id": "find-mersenne-rust",
      "name": "Find Mersenne Primes in rust",
      "description": "Finds the list of the first n Mersenne primes in Rust, or calculates a specific Mersenne prime for a given exponent parameter.",
      "tags": [
        "math",
        "benchmark"
      ]
    },
    {
      "id": "check-status-rust",
      "name": "Check calculation status in Rust",
      "description": "Returns 'ready' if no calculation is active, and 'not ready' if a calculation is live.",
      "tags": [
        "status",
        "health"
      ]
    }
  ],
  "capabilities": {},
  "defaultInputModes": [],
  "defaultOutputModes": []
}
Enter fullscreen mode Exit fullscreen mode

To check that that remote agent can support A2A:

xbill@penguin:~/a2a-multicloud/benchmark-rust-aws$ make a2a
Sending A2A status request to https://a2a-lightsail-rust-aws.6wpv8vensby5c.us-east-1.cs.amazonlightsail.com/...
{
  "id": 1,
  "jsonrpc": "2.0",
  "result": {
    "contextId": "status-context-id",
    "kind": "message",
    "messageId": "1245ea8c-e8d3-436a-8e27-0b25e90a8a42",
    "parts": [
      {
        "kind": "text",
        "text": "ready"
      }
    ],
    "role": "agent"
  }
}
Enter fullscreen mode Exit fullscreen mode

MCP Master/Orchestrator Agent (Local Rust)

The final agent follows a slightly different pattern. It provides a minimal agent with several functions (Tools/Skills).

To build-

make

xbill@penguin:~/a2a-multicloud/rust-master$ make lint
Linting the Rust project...
    Finished `dev` profile [unoptimized + debuginfo] target(s) in 0.09s
Enter fullscreen mode Exit fullscreen mode

To get the local endpoint:

xbill@penguin:~/a2a-multicloud/rust-master$ make endpoint
http://localhost:8100
Enter fullscreen mode Exit fullscreen mode

Then use the A2A inspector:

and check the A2A Agent card:

xbill@penguin:~/a2a-multicloud/rust-master$ make card
Fetching master agent card from http://localhost:8100...
{
  "name": "master_agent",
  "description": "Master A2A Agent that coordinates prime agents (including rust_agent, gcp_agent, aws_agent, and azure_agent) using model: gemini-2.5-flash.",
  "protocolVersion": "0.3.0",
  "version": "0.1.0",
  "url": "http://0.0.0.0:8100/",
  "skills": [
    {
      "id": "coordinate-agents",
      "name": "Coordinate sub-agents",
      "description": "Delegates requests to appropriate sub-agents for weather, time, events, prime calculations, and general greetings.",
      "tags": [
        "coordinator",
        "a2a"
      ]
    },
    {
      "id": "calculate-mersenne-prime",
      "name": "Calculate Mersenne Prime",
      "description": "Calculates Mersenne prime of n using connected sub-agents.",
      "tags": [
        "mersenne",
        "prime",
        "a2a"
      ]
    }
  ],
  "capabilities": {},
  "defaultInputModes": [],
  "defaultOutputModes": []
}
Enter fullscreen mode Exit fullscreen mode

and A2A status:

xbill@penguin:~/a2a-multicloud/rust-master$ make a2a
Sending A2A status request to http://localhost:8100/...
Enter fullscreen mode Exit fullscreen mode

Antigravity CLI with MCP Connection

The Master Agent starts a FastMCP server interface. This allows Antigravity CLI to be used as a MCP client:

xbill@penguin:~/a2a-multicloud/.agents$ more mcp_config
{
  "mcpServers": {
    "rust-master": {
      "command": "cargo",
      "args": ["run", "--manifest-path", "/home/xbill/a2a-multicloud/rust-master/Cargo.toml", "--quiet", "--", "--stdio"],
      "env": {
        "GOOGLE_GENAI_USE_VERTEXAI": "1",
        "GOOGLE_CLOUD_PROJECT": "aisprint-491218",
        "GOOGLE_CLOUD_LOCATION": "us-central1",
        "GENAI_MODEL": "gemini-2.5-flash",
        "MODEL_NAME": "gemini-2.5-flash"
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

The MCP tools can be reviewed:

MCP Servers

Plugins (~/.gemini/antigravity-cli/plugins)
   ✓ google-dev-knowledge Tools: search_documents, answer_query, get_documents
   ✓ rust-master Tools: ask_master_agent, calculate_mersenne_prime, check_agents_status
Enter fullscreen mode Exit fullscreen mode

The agent status can be checked with a MCP call to the Master Agent:

● rust-master/check_agents_status(Call check_agents_status MCP tool) (ctrl+o to expand)

  I ran the check_agents_status tool from the rust-master MCP server. The check succeeded, confirming that all sub-
  agents are healthy and online:

  • AWS (Remote): ✅ ONLINE (HTTP 200)
  • GCP (Remote): ✅ ONLINE (HTTP 200)
  • Azure (Remote): ✅ ONLINE (HTTP 200)
  • Rust (Local): ✅ ONLINE (HTTP 200)

  ### Summary of Work:

  1. Inspected Makefile & Scripts: Reviewed target definitions ( make status , make a2a ) and script configurations.
  2. Checked Endpoint Health: Verified the agent URL parameters and credentials from .env .
  3. Executed MCP Status Tool: Invoked the lazy-loaded MCP tool check_agents_status from the rust-master 
  coordinator, confirming all remote and local calculation sub-agents are online and ready to accept tasks.
Enter fullscreen mode Exit fullscreen mode

and the agent can be verified:

> ask_master_agent what do you do

● rust-master/ask_master_agent(Call ask_master_agent MCP tool) (ctrl+o to expand)

  I called the ask_master_agent tool with the query "what do you do".

  Here is the response from the Master Agent:

  │ "I am a Master A2A Agent. I can delegate tasks to specialized sub-agents (Rust, GCP, AWS, or Azure agents) to
  │ calculate Mersenne primes or check exponents."

  ### Summary of Work:

  1. Invoked the MCP Tool: Called ask_master_agent on the rust-master server with the prompt what do you do .
  2. Presented the Results: Relayed the master agent's self-description as a coordinator for distributed multi-cloud
  Mersenne prime calculations.
Enter fullscreen mode Exit fullscreen mode

So What is all this Doing?

All of the setup and configuration allows you to start debugging and visualizing Agent flows using the A2A protocol. The remote Agents haven been validated using the A2A Inspector. The Master Agent acts as a co-ordinator to delegate Mersenne Prime Calculations to remote A2A agents cross cloud.

Time to Starting Running some Numbers!

Now that the individual Agents have been validated — open antigravity and use the MCP tool to start the session:

> what does calculate_mersenne_prime do?

● Search(Grep search run_a2a_benchmark) (ctrl+o to expand)

  I will view the definition of run_a2a_benchmark (which implements calculate_mersenne_prime ) in rust-            
  master/src/main.rs .

● Read(/home/xbill/a2a-multicloud/rust-master/src/main.rs) (ctrl+o to expand)

  The calculate_mersenne_prime(n) tool executes a distributed benchmark to find and verify Mersenne primes (2ᵖ - 1)
  for exponents from 1 to n.

  ### How it Works:

  1. Assigns Tasks Round-Robin: It iterates through the exponents 1 ≤ i ≤ n and checks the readiness of the connected
  sub-agents (GCP, AWS, Azure, and Local) using a round-robin schedule.
  2. Delegates Computation: Once a ready agent is found, the master agent sends the exponent i to it via the A2A
  protocol. The sub-agent calculates whether 2ⁱ - 1 is prime.
  3. Aggregates Timings & Results: It measures the ready check time and the calculation time for each exponent.
  4. Saves Metrics: The benchmark results, including timing data and calculated primes, are compiled into a summary
  markdown table and persisted to disk (e.g., in benchmark_results.json).

  ### Summary of Work:

  1. Source Code Review: Inspected the main.rs function in rust-master/src/main.rs to extract the exact
  program flow and role of each agent.
  2. Explained Mechanics: Highlighted the round-robin task assignment, remote A2A calls, and metrics compilation
  features.
Enter fullscreen mode Exit fullscreen mode

Start the Benchmarking Session

Once all the agents have been started — the Master agent on port 8100 is connected over MCP with Antigravity CLI:


> run the calculate_mersenne_prime with 100

  I will execute the calculation of Mersenne primes up to exponent 100 by querying the Master Agent's HTTP MCP route.

● Bash(curl -s -X POST -H "Content-Type: application/json" -d '{"jsonrpc": "2.0", "id": 1, "method": ...) (ctrl+o to
expand)
  I will wait for the distributed calculation up to exponent 100 to complete.
  The distributed calculation up to exponent n = 100 has completed successfully!
  ### Distributed Mersenne Prime Generation Report
  • Timestamp: 2026-06-02T13:40:04-04:00 
  • Cumulative Duration: 55,625.57 ms (~55.6 seconds)
  #### Generated Mersenne Primes
  • 2² - 1 = 3 (via AWS Lightsail)
  • 2³ - 1 = 7 (via Azure Container Apps)
  • 2⁵ - 1 = 31 (via GCP Cloud Run)
  • 2⁷ - 1 = 127 (via Azure Container Apps)
  • 2¹³ - 1 = 8,191 (via GCP Cloud Run)
  • 2¹⁷ - 1 = 131,071 (via GCP Cloud Run)
  • 2¹⁹ - 1 = 524,287 (via Azure Container Apps)
  • 2³¹ - 1 = 2,147,483,647 (via Azure Container Apps)
  • 2⁶¹ - 1 = 2,305,843,009,213,693,951 (via GCP Cloud Run)
  • 2⁸⁹ - 1 = 618,970,019,642,690,137,449,562,111 (via GCP Cloud Run)
  ──────
  ### Coordinated Exponent Assignments
   Exponent | Assigned Agent | Ready Check Time | Calc Time | Result
  --------------------|------------------------|----------------------|----------------------|----------------------
   1 | GCP (Cloud Run) | 662.32 ms | 658.00 ms | not prime
   2 | AWS (Lightsail) | 418.86 ms | 117.14 ms | 3
   3 | Azure (Container Apps) | 330.12 ms | 366.37 ms | 7
   4 | Rust (Local) | 30.34 ms | 12.07 ms | not prime
   5 | GCP (Cloud Run) | 651.16 ms | 667.50 ms | 31
   6 | AWS (Lightsail) | 119.91 ms | 103.80 ms | not prime
   7 | Azure (Container Apps) | 337.03 ms | 338.40 ms | 127
   8 | Rust (Local) | 30.31 ms | 13.20 ms | not prime
   9 | GCP (Cloud Run) | 628.36 ms | 621.74 ms | not prime
   10 | AWS (Lightsail) | 108.49 ms | 113.11 ms | not prime
   11 | Azure (Container Apps) | 336.13 ms | 324.22 ms | not prime
   12 | Rust (Local) | 21.45 ms | 15.00 ms | not prime
   13 | GCP (Cloud Run) | 629.91 ms | 662.56 ms | 8191
   14 | AWS (Lightsail) | 111.40 ms | 119.59 ms | not prime
   15 | Azure (Container Apps) | 325.95 ms | 336.56 ms | not prime
   16 | Rust (Local) | 36.22 ms | 13.74 ms | not prime
   17 | GCP (Cloud Run) | 678.16 ms | 635.27 ms | 131071
   18 | AWS (Lightsail) | 111.02 ms | 115.05 ms | not prime
   19 | Azure (Container Apps) | 343.99 ms | 306.91 ms | 524287
   20 | Rust (Local) | 24.83 ms | 25.15 ms | not prime
   21 | GCP (Cloud Run) | 612.20 ms | 610.21 ms | not prime
   22 | AWS (Lightsail) | 103.64 ms | 121.31 ms | not prime
   23 | Azure (Container Apps) | 362.88 ms | 361.34 ms | not prime
   24 | Rust (Local) | 25.12 ms | 14.09 ms | not prime
   25 | GCP (Cloud Run) | 646.72 ms | 641.75 ms | not prime
   26 | AWS (Lightsail) | 98.67 ms | 109.74 ms | not prime
   27 | Azure (Container Apps) | 380.92 ms | 319.36 ms | not prime
   28 | Rust (Local) | 24.17 ms | 17.71 ms | not prime
   29 | GCP (Cloud Run) | 596.21 ms | 615.23 ms | not prime
   30 | AWS (Lightsail) | 110.32 ms | 102.89 ms | not prime
   31 | Azure (Container Apps) | 393.16 ms | 325.54 ms | 2147483647
   32 | Rust (Local) | 20.81 ms | 14.46 ms | not prime
   33 | GCP (Cloud Run) | 616.02 ms | 610.53 ms | not prime
   34 | AWS (Lightsail) | 98.09 ms | 113.49 ms | not prime
   35 | Azure (Container Apps) | 331.80 ms | 396.37 ms | not prime
   36 | Rust (Local) | 27.76 ms | 26.81 ms | not prime
   37 | GCP (Cloud Run) | 626.58 ms | 641.85 ms | not prime
   38 | AWS (Lightsail) | 108.36 ms | 110.98 ms | not prime
   39 | Azure (Container Apps) | 343.18 ms | 341.17 ms | not prime
   40 | Rust (Local) | 30.80 ms | 21.23 ms | not prime
   41 | GCP (Cloud Run) | 606.28 ms | 637.02 ms | not prime
   42 | AWS (Lightsail) | 106.25 ms | 116.14 ms | not prime
   43 | Azure (Container Apps) | 339.94 ms | 318.64 ms | not prime
   44 | Rust (Local) | 21.90 ms | 12.41 ms | not prime
   45 | GCP (Cloud Run) | 628.11 ms | 640.00 ms | not prime
   46 | AWS (Lightsail) | 100.35 ms | 112.84 ms | not prime
   47 | Azure (Container Apps) | 343.33 ms | 323.55 ms | not prime
   48 | Rust (Local) | 19.90 ms | 28.04 ms | not prime
   49 | GCP (Cloud Run) | 643.14 ms | 664.11 ms | not prime
   50 | AWS (Lightsail) | 113.88 ms | 103.43 ms | not prime
   51 | Azure (Container Apps) | 364.66 ms | 317.36 ms | not prime
   52 | Rust (Local) | 22.84 ms | 14.70 ms | not prime
   53 | GCP (Cloud Run) | 617.28 ms | 651.10 ms | not prime
   54 | AWS (Lightsail) | 103.60 ms | 114.45 ms | not prime
   55 | Azure (Container Apps) | 340.59 ms | 323.87 ms | not prime
   56 | Rust (Local) | 26.55 ms | 11.62 ms | not prime
   57 | GCP (Cloud Run) | 635.30 ms | 625.28 ms | not prime
   58 | AWS (Lightsail) | 106.18 ms | 117.21 ms | not prime
   59 | Azure (Container Apps) | 328.18 ms | 336.98 ms | not prime
   60 | Rust (Local) | 16.18 ms | 17.11 ms | not prime
   61 | GCP (Cloud Run) | 621.01 ms | 642.30 ms | 2305843009213693951
   62 | AWS (Lightsail) | 106.05 ms | 115.97 ms | not prime
   63 | Azure (Container Apps) | 322.90 ms | 346.61 ms | not prime
   64 | Rust (Local) | 25.02 ms | 13.26 ms | not prime
   65 | GCP (Cloud Run) | 610.11 ms | 627.14 ms | not prime
   66 | AWS (Lightsail) | 105.60 ms | 107.45 ms | not prime
   67 | Azure (Container Apps) | 337.70 ms | 323.28 ms | not prime
   68 | Rust (Local) | 15.38 ms | 20.14 ms | not prime
   69 | GCP (Cloud Run) | 625.35 ms | 643.10 ms | not prime
   70 | AWS (Lightsail) | 108.00 ms | 108.02 ms | not prime
   71 | Azure (Container Apps) | 323.01 ms | 331.95 ms | not prime
   72 | Rust (Local) | 31.18 ms | 28.03 ms | not prime
   73 | GCP (Cloud Run) | 647.98 ms | 627.10 ms | not prime
   74 | AWS (Lightsail) | 107.88 ms | 112.41 ms | not prime
   75 | Azure (Container Apps) | 351.71 ms | 340.42 ms | not prime
   76 | Rust (Local) | 21.00 ms | 25.80 ms | not prime
   77 | GCP (Cloud Run) | 624.00 ms | 647.06 ms | not prime
   78 | AWS (Lightsail) | 97.09 ms | 111.97 ms | not prime
   79 | Azure (Container Apps) | 337.90 ms | 346.54 ms | not prime
   80 | Rust (Local) | 24.42 ms | 13.95 ms | not prime
   81 | GCP (Cloud Run) | 624.80 ms | 631.11 ms | not prime
   82 | AWS (Lightsail) | 108.87 ms | 107.81 ms | not prime
   83 | Azure (Container Apps) | 332.57 ms | 322.07 ms | not prime
   84 | Rust (Local) | 19.99 ms | 16.02 ms | not prime
   85 | GCP (Cloud Run) | 622.47 ms | 651.63 ms | not prime
   86 | AWS (Lightsail) | 112.12 ms | 107.24 ms | not prime
   87 | Azure (Container Apps) | 332.57 ms | 340.05 ms | not prime
   88 | Rust (Local) | 24.79 ms | 29.77 ms | not prime
   89 | GCP (Cloud Run) | 646.21 ms | 667.60 ms | 618970019642690137449562111
   90 | AWS (Lightsail) | 102.15 ms | 113.82 ms | not prime
   91 | Azure (Container Apps) | 343.99 ms | 342.82 ms | not prime
   92 | Rust (Local) | 29.97 ms | 13.91 ms | not prime
   93 | GCP (Cloud Run) | 661.59 ms | 662.92 ms | not prime
   94 | AWS (Lightsail) | 112.50 ms | 112.80 ms | not prime
   95 | Azure (Container Apps) | 328.13 ms | 337.41 ms | not prime
   96 | Rust (Local) | 26.83 ms | 24.57 ms | not prime
   97 | GCP (Cloud Run) | 636.67 ms | 601.39 ms | not prime
   98 | AWS (Lightsail) | 105.62 ms | 112.77 ms | not prime
   99 | Azure (Container Apps) | 327.66 ms | 327.75 ms | not prime
   100 | Rust (Local) | 31.73 ms | 18.09 ms | not prime
  ──────
  ### Summary of Work:

  1. Invoked Calculation via HTTP Endpoint: Executed the calculate_mersenne_prime call with n = 100 against the
  Master Agent's HTTP POST interface.
  2. Aggregated Results: Processed the response containing two additional large Mersenne primes (2⁶¹ - 1 and 2⁸⁹ - 1)
  and the complete assignment distribution.
Enter fullscreen mode Exit fullscreen mode

Review Performance

The results of the benchmark are graphed:

Results

Antigravity CLI summarized the results of the Agent Benchmarking:

> summarize key results cross cloud
▸ Thought Process
  Analyzing Cross-Cloud Results
  Here is a cross-cloud summary of the key findings, focusing on architecture, network latency, authentication overhead,
  and calculation behaviors:

  ### 1. Latency & Network Transit Comparison

  The total round-trip time (RTT) for a single round of checks and calculation varies heavily by cloud service model:

  • AWS (Lightsail): ~232 ms RTT (120ms Ready Check + 112ms Calculation)
      • Strengths: Lightsail container instances behave very similarly to standard VMs, offering quick, persistent HTTP
      response times and low network hops.
  • Azure (Container Apps): ~679 ms RTT (341ms Ready Check + 338ms Calculation)
      • Characteristics: Azure Container Apps run inside a managed Kubernetes framework (K8s/Envoy ingress), adding
      slight container routing and warm-up latency.
  • GCP (Cloud Run): ~1,275 ms RTT (635ms Ready Check + 640ms Calculation)
      • Characteristics: Serverless scaling latency, combined with authentication checks, leads to the highest round-
      trip time.

  ──────
  ### 2. The Authentication Overhead (The "Why")

  There is a massive latency difference between GCP and the other clouds, caused by security configurations:

  • GCP Cloud Run: Requires secure OIDC identity authentication. For every single network call, the master agent calls 
  gcloud auth print-identity-token as a local subprocess. Spawning this CLI subprocess on every check and call
  introduces an extra ~300ms to ~500ms of local delay.
  • AWS & Azure: These sub-agents are deployed with publicly accessible HTTP endpoints that do not require OIDC tokens,
  avoiding any local CLI token generation overhead.
  ──────
  ### 3. Latency Symmetry (Network vs. Compute)

  Across all cloud providers, the Ready Check Time and Calculation Time are nearly identical:

  • AWS: ~120 ms vs ~112 ms
  • Azure: ~341 ms vs ~338 ms
  • GCP: ~635 ms vs ~640 ms

  Insight: This symmetry proves that network transport and handshake/auth overhead (ping and TLS/IAM checks) constitute
  over 95% of the latency. The actual math workload (calculating exponent primality) is extremely fast and has
  negligible latency impact at these exponent sizes.
  ──────
  ### 4. Cross-Cloud Protocol Success

  • The master agent successfully coordinated with all three major cloud providers (GCP, AWS, Azure) and the local
  runner.
  • Multi-cloud work distribution successfully validated Mersenne prime candidates, returning identical results
  regardless of which cloud hosting provider did the execution.

  ### Summary of Work:

  1. Analyzed Authentication Mechanics: Explored the main.rs subprocess overhead in the coordinator codebase.
  2. Evaluated Latency Patterns: Compared ready checks vs. calculation times to deduce that network transit and
  security handshakes dominate overall duration.
Enter fullscreen mode Exit fullscreen mode

GCP Remediation

To improve the GCP numbers- the Auth token is preset:

    export GCP_ID_TOKEN=$(gcloud auth print-identity-token)
Enter fullscreen mode Exit fullscreen mode

Then the benchmark is re-run. There is a slight variance on each run:

  Here is a summary of the performance and distribution metrics comparing the cloud providers and the local environment
  from the Mersenne prime calculation run (exponents 1 to 100):

  ### Overall Performance Metrics

   Environment | Tasks Ass… | Cold Start Ready… | Avg Warm Ready T… | Avg Calculation T… | Compute Perform…
  ------------------------|------------|-------------------|-------------------|--------------------|------------------
   Rust (Local) | 25 | 12.23 ms | 21.63 ms | 11.37 ms | #1 (Fastest)
   AWS (Lightsail) | 25 | 544.26 ms | 107.94 ms | 110.79 ms | #2
   GCP (Cloud Run) | 25 | 484.59 ms | 127.31 ms | 135.49 ms | #3
   Azure (Container Apps) | 25 | 20,711.57 ms | 348.39 ms | 354.26 ms | #4 (Slowest)
  ──────
  ### Key Findings

  1. Cold Starts & Overhead (Ready Time):
      • Azure (Container Apps) experienced a significant initial cold start latency of ~20.7 seconds on its very first
      assignment (exponent 3). Even once warm, its ready time overhead (~348 ms) remained roughly 3x slower than AWS
      and GCP.
      • AWS (Lightsail) and GCP (Cloud Run) showed excellent and comparable warm ready times of 108 ms and 127 ms
      respectively, with minor startup penalties (~500 ms) on their first task execution.
  2. Compute Performance (Calculation Time):
      • Local Rust is the fastest (~11.4 ms average) because it runs bare metal without any network traversal, agent
      orchestration, or cloud virtualization overhead.
      • AWS outperformed both GCP and Azure, running calculations in 110.79 ms average.
      • GCP followed closely behind AWS at 135.49 ms average.
      • Azure was the slowest cloud computing node, taking 354.26 ms average—roughly 3x slower than AWS and 2.6x slower
      than GCP.
  3. Mersenne Primes Discovered by Agent:
      • GCP (Cloud Run): Discovered 5 primes (Exponents: 5, 13, 17, 61, 89)
      • Azure (Container Apps): Discovered 4 primes (Exponents: 3, 7, 19, 31)
      • AWS (Lightsail): Discovered 1 prime (Exponent: 2)
      • Rust (Local): Discovered 0 primes (Since tasks were assigned in a round-robin format, the local agent was
      assigned exponents that are multiples of 4 (4,8,12,…), none of which are prime).
Enter fullscreen mode Exit fullscreen mode

Final Results

After adjusting the OIDC token for GCP- the benchmark was re-run:

Summary

The goal of the demo/article was to get basic Agents implemented across multiple clouds and benchmark the Agent performance finding Mersenne primes.

The key to the solution was the A2A (Agent to Agent) protocol. One local and three remote agents were presented — covering various combinations of deployment clouds.

Finally — a Master/Orchestrator agent was started to connect and delegate to the other agents via the A2A protocol. Antigravity CLI was used to connect to the Master Agent over MCP and execute the benchmarks.

Top comments (0)