Discuss: Recommended User Interface / Wire Protocol

[lukegalea]

Hi @StreetLamb,

I gotta say, I really love Rojak! Super impressive!

Have you given any thought to how to best expose a rojak backed agent/"agent swarm" to a front-end UI? I was thinking of writing a FastMCP to Rojak wrapper to expose it as a goal orchestration tool ("Give me a plan for this goal", "Make the following changes to the plan", "Okay, execute the plan".. watch for SSE or streamable events for progress) - so, exposing to clients both a compatible chat completion API as well as a streamable HTTP MCP server that wraps the rojak orchestrator. I think that'll work, but a few other possibilities seem promising and I'd love your feedback on approach ->

Something like langchain's agent-chat-ui could be enabled with a Langbridge API to Rojak bridge..

But most promising seems to be AG-UI - I think making Rojak "speak" AG-UI-Protocol would be very powerful, supporting all this fun stuff.

All that to say -> how have you been exposing rojak to the world? What do you recommend? And, if I'm to invest a bunch of time and energy into building a bridge, which one do you think would align best with the project?

[StreetLamb]

Hey @lukegalea, thanks for the kind words about Rojak. I really appreciate it!

To be honest, Rojak started as an exploration to integrate temporal's fault-tolerant capabilities into agentic workflows , and I haven’t focused on it much lately. I have mostly used it for small experiments and have yet to battle-test it in real-world scenarios, though I do think it has a lot of potential.

AG-UI sounds like an exciting addition, and I would be open if you are still willing to contribute :) A couple of questions: do you think the lack of response streaming in Rojak (a limitation from Temporal) would be a blocker for many of the features? What do you foresee would need to be added/updated on Rojak's end to make this work?

[lukegalea]

Hmm!! Great call out. I'm not sure it's a hard limitation of Temporal that we can't support response streaming. Couldn't we use signals from the activities back to the workflow itself to send chunks of partial completion along the way?