Making an autogenerating agent task scheduler (digraphs!) with @typesafeai Jev.
ddwr@JoymFL
SignalThe idea is to get Jev ranking the nodes all at once, build the digraph and make each track safe to execute parallely.
Making an autogenerating agent task scheduler (digraphs!) with @typesafeai Jev. The idea is to get Jev ranking the nodes all at once, build the digraph and make each track safe to execute parallely.
This would allow parallelism extraction out of tasks, without the user having to define the dependencies beforehand.
The goal of this tool is to be very aggresive in catching cycles, report to the human for correction and should fall back to creating dependencies rather than allow race conditions.
The idea is for this to sit between LLMs and agents.
So, essentially
User prompt for a complex task -> LLM breaks down into atomics -> This tool to create scheduling lists -> execute via agents
To test this, I got a sample workflow, that needs to be benchmarked and creates a report. github.com/Joymfl/dag-jev/blob/main/input.
I got a semblance of a result, but looking closer it isn't quite right:
for example step 3. build binary -> reads src/ is not dependant on step 0. clone repo
This is clearly wrong, and a race condition. There are a few more, if you look closer.
I don't think the tolerance is the issue, maybe the prompt? maybe it's too vague for it to understand. That would be next steps, along with actual benchmarks of latency, toks and cost compared to a baseline (haven't decided what the baseline comparison is going to be).
Why build this? The plan is to create safe scheduling lists for executors and get the most out of cost/execution ratio. The next step after this would be to build a work stealing scheduler on a gpu, and load balance these tasks across multiple small models on it (totally not an excuse to finally dive into gpu memory models)