Advanced example with multiple agents integrated with the orchestrator

[icaro-csmtec]

Hi there,

I'm currently exploring the library and was wondering if there are any advanced examples showcasing multiple agents working together within a more complex structure, integrated with the orchestrator.

Additionally, I'm looking for more detailed documentation on how to build a custom planner and how to customize the steps it generates.

Thanks in advance for your support and for the great work on the project!

[saqadri]

Hi there,

I'm currently exploring the library and was wondering if there are any advanced examples showcasing multiple agents working together within a more complex structure, integrated with the orchestrator.

Additionally, I'm looking for more detailed documentation on how to build a custom planner and how to customize the steps it generates.

Thanks in advance for your support and for the great work on the project!

@icaro-csmtec Thanks for this question! We currently have a basic example for the orchestrator here: https://github.com/lastmile-ai/mcp-agent/tree/main/examples/workflow_orchestrator_worker.

To override the planner, you can specify a planner on the orchestrator class, which can be an Agent or AugmentedLLM of its own.

I can certainly work on providing a more advanced example, and to best support you, I would love some more details about your usecase. You can describe that here or we can chat on discord (https://lmai.link/discord/mcp-agent or message me directly @saqadri).

[icaro-csmtec]

I’ve been experimenting with the library and started from one of the provided examples, adapting it to fit a more specific use case using a custom agent.

In my scenario, each agent can take six or more specific parameters. For example:

Prompt (Book Analysis Agent):

You are a literary analysis expert. Your task is to analyze any book based on the following six specific parameters:

1. Summary – Provide a clear and concise summary of the main events.
2. Central Themes – Identify and explore the main themes of the book.
3. Character Development – Analyze the main characters and how they evolve.
4. Author's Style – Describe the writing style, including language, structure, and techniques.
5. Historical and Cultural Context – Relate the book to the historical and cultural context in which it was written or is set.
6. Impact and Relevance – Evaluate the book’s impact on literature or society and its current relevance.

This analysis should be deep, well-founded, and suitable for students, researchers, or curious readers.

I also have other agents that review the outputs from the first one, focusing on one or more of the above parameters. Finally, I have a third agent that compares all outputs based on a database of PDFs.

The issue I'm facing is that when I test this setup—especially using interactive mode instead of plan_type=full—I often get an error where the system fails to recognize the agent names, or mismatches them.

Because of this, I’d like to better understand:

• How can I control the steps of the plan to preserve the reference to my predefined agents?
• Is there a recommended way to ensure the orchestrator always focuses on the agents I've configured?
• When building the task, I provide detailed specifications and an initial value to be analyzed—how can I make sure this structure is properly respected in planning?

Thanks again for the support! I’d appreciate any guidance or examples you could share related to more advanced usage like this.

[saqadri]

Thank you for the indepth answer! Acknowledging receipt, will revert back after discussing internally with the team