This post summarizes nicely many what we have been building to enable coordination across multiple agents, tools, etc in GRID. This is an important problem to solve for open source monetization: we need to figure out effective ways of coordinating many open source artifacts into use AGI workflows automatically. Here is what we have done so far: First, at the outset, we found that LLMs lack knowledge of complex workflows which human experts use. So we curated those from experts for grounding (in future, this will be open for others to contribute) Next, we developed a technique for breaking queries into simpler tasks recursively. This deep research approach, termed ROMA, will be released as an open source library soon. Recently, @ManusAI_HQ used a related approach which they called wide search (although it is not open source, and no details were provided) Finally, we have been experimenting with many new “simple tasks” that constitute the leaf nodes in our flows. A fun one mentioned in this tweet thread is using multiple models composed using token-level routing, an interesting idea for combining the capabilities of two models. You will be able to experience queries answered in this manner on sentient chat over the next few months. A lot of ideas are going into prod for us!