a2a by jmoreno45 · Pull Request #17 · llnl/mada

jmoreno45 · 2026-07-30T18:18:31Z

Adds A2A capability with some example agents.

python examples/a2a_table_mcp_server.py --port 9101 &
python examples/a2a_average_mcp_server.py --port 9102 &
python examples/a2a_langchain_agent.py --port 9111 --config configs/example_a2a_agents.json --mcp-url http://localhost:9101/mcp &
python examples/a2a_google_adk_agent.py --port 9112 --config configs/example_a2a_agents.json --mcp-url http://localhost:9102/mcp &
mada-cli configs/example_a2a_agents.json
mada-gradio /g/g20/moreno45/Projects/GITHUB-MADA/mada/configs/example_a2a_agents.json

Can also be exposed as an A2A agent

Run MADA in A2A mode on the same port advertised in the card:

mada-a2a --port 9120 /g/g20/moreno45/Projects/GITHUB-MADA/mada/configs/example_a2a_agents.json

Other A2A agents can discover it at:

http://localhost:9120/.well-known/agent-card.json

If another machine/agent needs to reach it, use --public-url with the externally reachable URL:

mada-a2a --port 9120 --public-url http://hostname:9120/ /g/g20/moreno45/Projects/GITHUB-MADA/mada/configs/example_a2a_agents.json

CLI

Initializing agents and MCP servers...
Status: Connection Successful: Orchestrator initialized with 0 MCP Servers and 2 remote A2A agents and 3 agents
Orchestration mode: agent-as-tool
Model: gpt-5.4 from livai

Available tools:
   • A2A: LangChainAgent - Table reader specialist that can load and display a small built-in CSV table by calling its MCP table-reader tool.
   • A2A: GoogleADKAgent - Column average specialist that can compute averages for numeric columns in a small built-in CSV table by calling its MCP average tool.

Chat with the agents (type 'quit' to exit)
--------------------------------------------------
 
You: load table

Agents:
--------------------
[task-1] Started in background.


[task-1] Completed:

[Calling: call_langchainagent]
Loaded table:

| experiment | temperature_c | pressure_kpa |
|---|---:|---:|
| alpha | 21.2 | 101.3 |
| beta | 24.8 | 99.8 |
| gamma | 19.6 | 103.1 |
| delta | 22.4 | 100.6 |

 
You: average table

Agents:
--------------------
[task-2] Started in background.


[task-2] Completed:

[Calling: call_googleadkagent]
Column averages:

- temperature_c: 22.00
- pressure_kpa: 101.20

 
You:

Gradio

bgunnar5 · 2026-07-30T23:40:21Z

If another machine/agent needs to reach it, use --public-url with the externally reachable URL:

mada-a2a --port 9120 --public-url http://hostname:9120/ /g/g20/moreno45/Projects/GITHUB-MADA/mada/configs/example_a2a_agents.json

Haven't looked at any of the code here yet but this sounds like it would be a perfect use-case for Wormhole.

bgunnar5

Nice work, this will definitely help to expand MADA outreach and usability in the future. In addition to the comments I left, here are some additional requests:

Add documentation on the mada-a2a command.
Can we group a2a examples in a subfolder under examples/?
If we don't implement the A2AAgent from Microsoft Agent Framework, then we should add tests for the new a2a_client.py file.

@Mamorri there may be some modifications you need to add to your magentic implementation to support A2A.

Don't need you to do this but I'm putting this here to remind myself to create a bunch of tutorials for MADA. We should definitely use the example scripts you put together for a full A2A tutorial. It would also be a good idea to have a tutorial for starting URSA and connecting it via A2A with MADA. This could lead to a concrete agent card for MADA x URSA collaboration that we could keep in our repo as well.

bgunnar5 · 2026-08-03T19:24:59Z

+"""
+Agent-to-Agent HTTP interface for MADA Orchestrator.
+
+This module exposes the configured MADA planning agent as an A2A-compatible
+JSON-RPC service. The MADA agent card is available under the standard
+`/.well-known/agent-card.json` path.
+
+This is the server-side A2A entry point: use it when another A2A client or
+agent needs to discover MADA and send work to MADA. The client-side support
+for MADA calling other A2A agents lives in `mada.core.a2a_client` and is wired
+through the `a2a_agents` configuration block.
+"""


so this is only starting the planning agent as an A2A-compatible service? Would it make more sense to do something like mada-a2a config.json and then every agent defined in the config file is started as an A2A-compatible service?

This should start the orchestrator which then has access to all the other agents listed in the config, including mcp tools. So its a central entry point for mada. Up to you, what do you prefer?

await orchestrator.initialize_orchestrator( self.config.agents, self.config.mcp_servers, getattr(self.config, "a2a_agents", {}), )

This is probably fine for now. Just not sure how this will behave with @Mamorri's orchestration updates, specifically for sequential which likely won't need a planning agent.

I went ahead and merged these changes into my branch, not too many conflicts directly. Functionality wise, there may be some overlap. It looks like for magentic there was a small breakage just in how the new arguments were set up, but it should be a fairly small adjustment. However, once restored A2A should complement the dynamic orchestration functionality.

For sequential, if remote A2A agents are only implemented as planner tools, sequential would need its own way to include them as steps. My best thought here would be a “participant/callable agent” interface that local specialists and remote A2A agents both implement. Then agent-as-tool, magentic, and sequential could each decide how to use the same pool.

For now I think it should be fine.

jmoreno45 · 2026-08-04T23:26:00Z

Nice work, this will definitely help to expand MADA outreach and usability in the future. In addition to the comments I left, here are some additional requests:
* [ ]  Add documentation on the `mada-a2a` command.

* [ ]  Can we group a2a examples in a subfolder under `examples/`?

* [ ]  If we don't implement the `A2AAgent` from Microsoft Agent Framework, then we should add tests for the new a2a_client.py file.
@Mamorri there may be some modifications you need to add to your magentic implementation to support A2A.

Don't need you to do this but I'm putting this here to remind myself to create a bunch of tutorials for MADA. We should definitely use the example scripts you put together for a full A2A tutorial. It would also be a good idea to have a tutorial for starting URSA and connecting it via A2A with MADA. This could lead to a concrete agent card for MADA x URSA collaboration that we could keep in our repo as well.

Is this the documentation you are referring to towards the bottom or would you like additional information? https://github.com/llnl/mada/pull/17/changes#diff-058b5883ae0b9c88b6008b2d3be1b7a4b9d8497e83b7e1dfd8475e953ec6178cR328
I have grouped a2a in examples/a2a
I have updated to use A2AAgent and a2a as well

bgunnar5 · 2026-08-05T14:43:56Z

Is this the documentation you are referring to towards the bottom or would you like additional information? https://github.com/llnl/mada/pull/17/changes#diff-058b5883ae0b9c88b6008b2d3be1b7a4b9d8497e83b7e1dfd8475e953ec6178cR328

@jmoreno45 I think I missed that on the first review, that's fine.

bgunnar5

This LG2M but I'd like to hear Michael's thoughts on some of my comments that I tagged him in.

Mamorri · 2026-08-06T16:08:28Z

+            card = self._a2a_agent_cards.get(agent_name, {})
+            description = self._remote_a2a_description(
+                card["description"],
+                card,


After merging my changes it looks like this presented a potential issue:
remote A2A cards are indexed withcard["description"]. If a remote card omits description or serializes differently, orchestrator initialization fails with KeyError. I think agent_as_tool_strategy.py and other orchestration strategies would require a method to parse the remote description:

description = orchestrator._remote_a2a_base_description(agent_name, card) labels.append(f"A2A: {agent_name} - {description}")

I can go ahead and add this to my commits, if it makes more sense than fixing in this merge

Yes, please add it to your commits.

Jorge Moreno and others added 6 commits July 30, 2026 11:16

a2a

0725325

self a2a card

fb736a3

docs

ccbb72e

changelog

c2ac55d

updating docstrings

95caca9

adding docstrings

6679b72

jmoreno45 requested a review from bgunnar5 July 30, 2026 22:15

updating sample agents

17bc549

bgunnar5 requested changes Aug 3, 2026

View reviewed changes

Jorge Moreno added 2 commits August 3, 2026 15:25

github comments

1878a30

using agenta2a

46415d9

bgunnar5 marked this pull request as ready for review August 5, 2026 14:44

bgunnar5 approved these changes Aug 5, 2026

View reviewed changes

Mamorri reviewed Aug 6, 2026

View reviewed changes

merging in develop

64da59c

jmoreno45 merged commit 64943f0 into develop Aug 7, 2026
53 of 54 checks passed

Uh oh!

Conversation

jmoreno45 commented Jul 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Adds A2A capability with some example agents.

Can also be exposed as an A2A agent

CLI

Gradio

Uh oh!

bgunnar5 commented Jul 30, 2026

Uh oh!

bgunnar5 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bgunnar5 Aug 3, 2026

Choose a reason for hiding this comment

Uh oh!

jmoreno45 Aug 4, 2026

Choose a reason for hiding this comment

Uh oh!

bgunnar5 Aug 5, 2026

Choose a reason for hiding this comment

Uh oh!

Mamorri Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jmoreno45 commented Aug 4, 2026

Uh oh!

bgunnar5 commented Aug 5, 2026

Uh oh!

bgunnar5 left a comment

Choose a reason for hiding this comment

Uh oh!

Mamorri Aug 6, 2026

Choose a reason for hiding this comment

Uh oh!

jmoreno45 Aug 7, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

jmoreno45 commented Jul 30, 2026 •

edited

Loading