Architectural decitions regarding manual topology tests

[guillermo-delrio]

There's been some discussions recently on how to proceed with the manual topology tests. Preliminary decisions have been taken, but I look forward to reading the feedback from other participants.

All these actions made are meant to aid on quickly drafting up manual tests including a variety of topologies, prompts, and functions.

Georgia and Chad from the ACE team had some very good ideas based on their experience preparing the ACE demos.

We want to take some time to bring the work done on ACE into HAAS, to avoid working twice on the same issues.

In that sense Chad will be looking into moving the system to docker containers, with basic communication using RabbitMQ. We recognize that this moves away from the nice "hackability" of a simple Python script, but looking forward a few weeks in the development it seems this is where we're going to get anyway, so it might be better to bite the bullet now.

Georgia has started looking into the "raise hand" idea to prevent an agent from generating further content before it has a chance to catch up with the conversation.

I am inclined to work on the implementation of a self-prompting "give me more time to think" function call. To help improve the quality of the messages pushed to the shared channels.

Lastly, the need for a good conversation visualisation and debugging system has been identified. This may initially take the form of a simple UI that displays each thread messages in order for each agent; but in the future it may be expanded to show messages per channel, connections between agents, and function calls.

All that work would be done in parallel to the self-sufficiency work than by other collaborators. To avoid interrupting those efforts we will be working on a separate folder structure. After it is pushed and tested we might reconvene to check if this approach was successful and may be extended to the whole project.

I hope this proposal makes sense. I’m happy to discuss any of the points to try to reach consensus.