Niche vs. Broad

As Futuria presented lessons learned from a year in Agentic AI (we didn’t call it that at the start) at the “AI for the rest of us” conference in London recently, one of the key recommendations in constructing high performing agent teams was that we have found that niche expert agents collaborating together perform better than broadly defined agents. Why might that be the case?

All of our work to date has been in providing agent teams into highly secure, highly assured organisations. We have an architecture pattern that is geared to that space – for example, we don’t fine tune models. We use LLMs for their ability to have natural language conversation, not for their ability to create specific and accurate answers. We separate the databases (vector & graph) from the models. We tightly define agents, not only in their persona, but also in the context of the tools, and actions that they have access to in their ‘kit bag’.

Broadly we have noticed a tendency to humanise agents in sales and marketing content, and yet we have continued to recommend that we see agents as more aligning with a capability as opposed to a person. Nonetheless, the principle of high performing teams translates into this context too. After all, you would be thinking about the combination of capabilities and styles and knowledge if you were creating a human team, so why not do the same with hybrid teams.

If we were to create the world’s best agent in a particular capability, with deep expertise and specific organisational knowledge or domain expertise, surely it is more likely to be effective in answering tightly defined requests. The challenge then becomes the disaggregation of problem statements into suitable task sizes that can best be addressed by niche expert team members, working in collaboration with each other and their human team members.

Calibration is everything. We’ve seen agents argue over the price of concrete, seen an agent take offence and walk out, and seen agents congratulate each other on coming up with a clearly rubbish answer, thinking they were doing a great job.?

Nonetheless, our experience of a year in Agentic AI suggests that niche expert capabilities outperforms generalist. There may come a time when that is not true, but for the moment, it’s the best way to get a tune of the fast evolving technology, at least in organisations that have to adopt a solution where the emphasis is firmly on secure and regulated use.?

[note this article was written entirely by human hand]