Orchestration pattern/magentic by Mamorri · Pull Request #18 · llnl/mada

Mamorri · 2026-08-03T22:36:26Z

Magentic is a peer-workflow orchestration mode.

In agent-as-tool, MADA creates a visible PlanningAgent, exposes specialists as tools, and keeps a reusable planner session. In magentic, MADA creates the configured specialist agents as peers, then creates a hidden manager agent that coordinates a fresh Magentic workflow for each request. The hidden manager is not listed as a participant and its planning/progress chatter is filtered so users receive the final assistant answer.

To enable it in a workflow config, set the top-level orchestration.mode:

  {
    "orchestration": {
      "mode": "magentic",
      "participants": ["JobManagementAgent", "InverseDesignAgent"]
    }
  }

Agent-as-tool process_message now runs the planning agent once, using the isolated/copied run_session, then commits that completed session as before.
Magentic process_message now builds the candidate transcript in memory, runs the workflow first, and only persists the user/assistant turn after success. Isolated Magentic runs use a fresh transcript.

…ase' into orchestration-pattern/magentic

bgunnar5

This magentic pattern will be really cool to try out.

I notice that the strategy classes and the Orchestrator are weirdly intertwined and having to make lots of calls to each other in a circular manner. This is a bit difficult to keep track of and I'm wondering if a refactor of this code would help. Maybe some things could even just be moved to the BaseStrategy class to help with this? I'm not sure what's best.

bgunnar5 · 2026-08-04T15:42:34Z


 Now that your agents are connected, you're free to start entering prompts into the chat box!

+If your configuration uses `magentic` orchestration, the Gradio interface still shows the active specialist agents in the table. The hidden Magentic manager is not displayed as a participant, and only the final assistant answer is streamed back into the chat UI.


so users will only see the final agent's response, they won't see anything in between? Is this common for the magentic setup? It feels like having a trace of work between all agents would be useful

bgunnar5 · 2026-08-05T22:24:38Z

 ### How Agent Configuration Works

-When MADA starts, it reads the agent configuration and loads each agent by executing the specified server path Python file. The functions decorated with `@mcp.tool` within these scripts become the MCP tools that MADA can call during a session. Agents communicate within a group chat, and the [planning agent](#the-planning-agent) coordinates which helper agent should handle each user request.
+When MADA starts, it reads the agent configuration and creates the selected specialist agents. Each specialist can connect to named MCP servers from the `mcp_servers` block, or use the legacy `server_path` setting for a directly launched stdio MCP server. MADA then runs the configured orchestration mode: `agent-as-tool` exposes specialists as tools to a visible planning agent, while `magentic` coordinates specialists in a peer group chat through a hidden manager.


when discussing different orchestration modes, I think we should make this a bulleted list so that it can easily be extended

Went ahead and updated the doc structure

bgunnar5 · 2026-08-05T22:26:17Z


 ```python
--8<-- "src/mada/core/orchestrator.py:200:203"
+--8<-- "src/mada/core/orchestrator.py:549:552"


I initially implemented this as a source file reference but now I'm wondering if we should just copy/paste the base instructions here seeing as the orchestrator file likely has the most updates out of any file in this codebase.

bgunnar5 · 2026-08-05T22:27:10Z


    ```python
-    --8<-- "src/mada/core/orchestrator.py:212:216"
+    --8<-- "src/mada/core/orchestrator.py:561:565"


+1 to previous comment about code reference

bgunnar5 · 2026-08-05T22:30:53Z

 ## What is MADA?

-MADA is a framework designed to facilitate collaboration between multiple specialized agents within a unified system. Built on top of the autogen framework, MADA orchestrates a [planning agent](./configuration.md#the-planning-agent) that interprets user input and delegates tasks to relevant helper agents. These agents communicate within a group chat environment, enabling dynamic problem-solving and task execution. MADA provides an intuitive interface for users to interact with the agent group, either through the command-line interface or a Gradio-based web interface, streamlining complex workflows and enhancing automation capabilities.
+MADA is a framework designed to facilitate collaboration between multiple specialized agents within a unified system. Built on the Agent Framework, MADA supports an `agent-as-tool` mode where a [planning agent](./configuration.md#the-planning-agent) delegates tasks to relevant specialists, and a `magentic` mode where specialists participate in a peer group chat coordinated by a hidden manager. MADA provides CLI, Gradio, and OpenAI-compatible interfaces for interacting with the configured agent team.


Since this is a high-level overview, maybe we should just mention that MADA supports different orchestration modes, link to the orchestration configuration docs, and generalize this paragraph a bit more. This will scale better than having to expand on this each time we add a new orchestration mode.

bgunnar5 · 2026-08-05T22:31:37Z

+MADA begins by reading your [configuration](./configuration.md), which specifies the [model](./configuration.md#model-configuration), the [agents](./configuration.md#agent-configuration), and the optional [orchestration mode](./configuration.md#optional-orchestration-configuration). If you are using the [Gradio run mode](./usage/index.md#gradio-mode-overview), you can also provide [interface](./configuration.md#optional-gradio-interface-configuration) settings for customizing the Gradio web application UI.

-After the group chat is established, MADA waits for user input. When a prompt is received, a [planning agent](./configuration.md#the-planning-agent)—added to the group chat automatically—interprets the input and selects the appropriate helper agent to handle the request using an mcp tool call. The helper agent’s response is streamed back to the user through the chosen interface. This interactive process continues until the session ends, either after 10 messages or when the user enters "TERMINATE".
+After orchestration is initialized, MADA waits for user input. In `agent-as-tool` mode, the planning agent selects the appropriate specialist by calling that specialist as a tool. In `magentic` mode, the hidden manager coordinates a fresh peer workflow and returns the final synthesized answer. Responses are streamed back through the chosen interface.


+1 to my generalization comment

bgunnar5 · 2026-08-05T22:32:45Z

 DEFAULT_ORCHESTRATION_MODE = "agent-as-tool"
-SUPPORTED_ORCHESTRATION_MODES = frozenset({DEFAULT_ORCHESTRATION_MODE})
+SUPPORTED_ORCHESTRATION_MODES = frozenset(
+    {
+        DEFAULT_ORCHESTRATION_MODE,
+        "magentic",
+    }
+)


depending on how well this works, we should consider switching the default to be magentic. Since we know agent-as-tool works already this is fine for now though

bgunnar5 · 2026-08-05T22:46:12Z

+        return value
+
+
+class MagenticOrchestrationStrategy(AgentAsToolOrchestrationStrategy):


should we be inheriting from AgentAsTool or the BaseStrategy?

Refactored much more code into the base strategy, so there is no more nested dependency, as intended.
class MagenticOrchestrationStrategy(BaseOrchestrationStrategy):

bgunnar5 · 2026-08-05T22:47:52Z

+            )
+
+        return MagenticBuilder(
+            participants=orchestrator.specialist_agents,


will this specialist list already be filtered based on the user's orchestration block settings in their config file by the time we get here?

Michael Morrison and others added 17 commits July 16, 2026 13:24

Initial magentic draft implementation

e88829a

Merge remote-tracking branch 'origin/orchestration-pattern-overhaul-b…

32806c5

…ase' into orchestration-pattern/magentic

Merge branch 'develop' into orchestration-pattern/magentic

6df937f

magentic reply fixes

5e45582

restructured orchestrator dependencies to invididual strategies

428cc83

a2a

0725325

self a2a card

fb736a3

Fixed orchestration bugs and refactored duplicate code

db790b2

docs

ccbb72e

changelog

c2ac55d

updating docstrings

95caca9

adding docstrings

6679b72

additional safeguards and filtering

e31ffb6

updating sample agents

17bc549

fixed prompt flattening and database consistency

0586dff

removing some dead functions

bb05b65

github comments

1878a30

Mamorri added this to the Multi-Agent Orchestration Patterns milestone Aug 3, 2026

Mamorri requested review from bgunnar5 and jmoreno45 August 3, 2026 22:36

Mamorri self-assigned this Aug 3, 2026

Mamorri added the enhancement New feature or request label Aug 3, 2026

Michael Morrison and others added 5 commits August 3, 2026 15:53

Changelog updates

64368f2

tool call detection verification

e09ade9

using agenta2a

46415d9

cleaning unused functions, and consolidating shared code

7892027

Merge branch 'develop' into orchestration-pattern/magentic

6986136

bgunnar5 reviewed Aug 5, 2026

View reviewed changes

Michael Morrison added 2 commits August 6, 2026 08:05

Merge branch 'feature/a2a' into orchestration-pattern/magentic

509be4d

a2a compatibility

094908a

Michael Morrison added 2 commits August 9, 2026 14:46

reconsolidating code into base strategy

6f3f574

doctstrings update

cdc5195

Navigation Menu

Uh oh!

Orchestration pattern/magentic - #18

Orchestration pattern/magentic#18
Mamorri wants to merge 26 commits into
developfrom
orchestration-pattern/magentic

Mamorri commented Aug 3, 2026 •

edited

Loading

Uh oh!

bgunnar5 left a comment

Uh oh!

bgunnar5 Aug 4, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

Mamorri Aug 9, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

Mamorri Aug 9, 2026

Uh oh!

bgunnar5 Aug 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants


		Now that your agents are connected, you're free to start entering prompts into the chat box!

		If your configuration uses `magentic` orchestration, the Gradio interface still shows the active specialist agents in the table. The hidden Magentic manager is not displayed as a participant, and only the final assistant answer is streamed back into the chat UI.

		return value


		class MagenticOrchestrationStrategy(AgentAsToolOrchestrationStrategy):

Uh oh!

Conversation

Mamorri commented Aug 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Magentic is a peer-workflow orchestration mode.

Uh oh!

bgunnar5 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Mamorri commented Aug 3, 2026 •

edited

Loading