Orchestrator termination

[mattwg]

Thanks to everyone who has helped create this mcp based agent framework. I think it makes a lot of sense.

I have been experimenting with the Orchestrator pattern - and providing it with file system tools and web search tools. I have a few questions:

• I am often seeing the main task get successfully completed but the orchestrator does not realize it is done and it kicks off other tasks. Has anyone else observed this and any tips for how to reduce the likelihood this happens. For example I may ask a questions about a document and it is clear the agents have answered it - only for the orchestrator to ignore that we are done. I was using o3-mini and I wondered if maybe that model was not sophisticated enough to be the orchestrator?

• My console output in the terminal is printing like this:

With the useful bits displayed in amongst all the Finished / Running updates

This makes reading it real hard - any tips to get a more easily readable output for debugging and exploring the agents decisions?

Appreciate any pointers.

Cheers,
Matt

[saqadri]

Hi @mattwg thank you for supporting the framework! Yes I have observed this too with o3-mini. Interesting this doesn't happen with gpt-4o. Can you please create an issue? I am working on improving the orchestrator and can resolve this.

[mattwg]

Thanks @saqadri - I'll swap it out and report back! If you give me some pointers I am happy to help contribute to the orchestrator. I assume this is a prompt/model related issue - trying to find a prompt that works for all models might be challenging...