The Emperor’s New AI agents

The enterprise AI vendor landscape is full of sophisticated-sounding buzzwords: coordinated orchestrators, hierarchical agent networks, autonomous swarms making decisions in real time. The announcements are impressive. The outcomes… are not.

A February 2026 study by the National Bureau of Economic Research, surveying nearly 6,000 executives across four countries, found that 89% of firms reported zero change in productivity from AI ( NBER Working Paper #34836). The obvious gap between what CEOs say and what is actually engineered to work is difficult to ignore.

The gap is not a technology problem. We believe it is an implementation sequence problem. And the sequence most enterprises follow is backwards. Let me explain.

The emperor's new agents #

The gap between the announcement of turning agentic and achieving real outcomes recalls the well-known story of the Emperor's New Clothes. Everyone insists that multi-agent systems have arrived, but they're pointing at a naked man in the street. The reality is, that pilots and POCs (proofs-of-concepts) outweigh productive Agentic AI systems by far. Aside from a few failed chatbots for customer service that can't help achieve much of anything, genuinely productive multi-agent AI remain elusive. And no one wants to admit that companies are adopting agents that fail to work.

Having build and shipped several systems that do reliably work in production, we at Vstorm see the reasons that others fail does not come from a single factor. But in this post we'll cover just one, one which impacts the ability to ship productive agents big time. Enterprises deploy from the wrong end. Instead of building agentic AI capabilities bottom-up, they go top-down.

Why top-down orchestration fails #

We have seen many overly ambitions projects with agentic architecture all laid out before any single RAG component got deployed. We have seen ill conceived workflows where 'agentic components' were planned on top of human-driven processes without rethinking those processes from ground up. Such cases are the digital equivalent of skeuomorphism: replicating the old approach or obsolete design just because if feels a too familiar.

But above all, the narrative for multi-agent orchestration is out of sync with where the technologies are. Top-down planned multi-agent systems look good on paper, but don't yet work as many little points of failure derail such initiatives. Partly because of a lack of experience, partly due to the immaturity of the tech.

A big problem is that in a pipeline of ten agents, even 95% per-agent accuracy produces approximately 60% accuracy overall, because each agent treats its predecessor's output as ground truth. So these errors do not cancel. They compound.

There is an alternative, however.

The bottom-up alternative: one working agent and beyond #

All of the projects we successfully shipped started simple, as a single agent with very high accuracy and reliability, deployed against one workflow, with one measurable success criterion. Such an agent, once it actually does what it should, deployed to production does the magic of giving return on investment. And the ROI from Agentic AI is fuel for expansion for our middle-market customers.

The next step, however, isn't what you think; it's not immediately going multi-agent. It's rather expanding the single agent, adding tools, extending its scope, and taking it to its limit, so it can take on new and more valuable roles. A single all-capable agent, operating reliably in production, is a stronger foundation than ten agents coordinating on paper. Our advice:

Start small. Do not just jump head-first into all the complexity of orchestrating a multi-agent system. Instead, find a way to leverage what a single one can do, take it to the limit and then, only then, consider splitting it into a more capable, multi-agent system. – Wojciech Achtelik, PhD and AI Engineer Lead of Vstorm