Agent-oriented systems—which include software that accepts instruction and then takes action on your behalf—have historically underperformed. Every ten years or so, we experience yet another wave of agent hype.

Oddly, most software and services providers don’t seem to remember or take the most important lessons away from the previous waves of hype.

In 1995, Gartner had “intelligent agents” at the peak of inflated expectations. In 1997, Microsoft introduced Clippy.

In 2015, when I was researching emerging tech at PwC,  we had another dose of inflated expectations, which prompted me to dig up the 1995 Gartner hype cycle chart shown here.

Gartner, 2015

In 2025, the “inflated expectations” have returned with a vengeance. Given that we have just barely entered the phase of contextual computing that John Launchbury of DARPA called the Third Wave back in the 2010s, most aren’t aware of the best practices needed to build and connect contexts. 

Technologists want to use agents in a much bigger way, and businesspeople clearly also understand the value of an agent orchestration capability, as evidenced by the number of vendors now offering products they claim will help with orchestration.  

But very little investment goes to serious contextualization and the FAIR (findable, accessible, interoperable and reusable) data and flexible data models needed for reliable, scalable agent orchestration. Investors, for one thing, don’t invest in many seriously capable systems to begin with. For another, the systems they do invest in lack cohesion that scales.

What is “Cohesion”? 

Ravi Patel in a Medium post describes quantifiable “structural cohesion” from a software engineering perspective this way:

“Data Dependency: Evaluates how data is shared among methods within a class. High cohesion is indicated by frequent data sharing among methods.

Graph-Based Metrics: Uses graphs where nodes represent methods and edges represent shared data or interactions. Cohesion is measured by the density of connections within the module.”

From this perspective alone, traditional software often succeeds at small scale, but on its own at larger scale, it often fails to serve bigger, more universal purposes and more sizeable communities of users. Most software, after all, relies on relational or document databases.Those databases need lots of ancillary, external help to scale and serve larger interconnected communities of users.

Cohesion from a software perspective, in short, is the ability to share data in an interactive way.

When Graphs Become Mandatory for Cohesion at Scale

Graphs come into play much more often when the purpose requires large numbers of interconnections between domains, not to mention different classes, entities and flexibly modeled relationships throughout. At some point, architects are modeling whole business units, industries and interactions between organizations at supply chain scale. 

This is the reason that semantics technology professionals rely on graph databases and semantic metadata tooling to manage knowledge representation–so scaling becomes possible while maintaining accuracy and clarity. And it’s the reason these professionals spend many hours contributing to standards that facilitate interoperation and reusability. The following illustration uses a biomedical example to  place classes, domains and entities in their respective contexts, so you get a sense of how knowledge modeling enables the ability to scale out.

 Image generated by Claude.ai, September 2025

Not Just Big Data, but Big Knowledge

So many people are starting to understand that what’s essential on the input side of AI for agents to perform reliably, doing our bidding in a helpful, safe and efficient way at scale. The missing component is actually the right knowledge. And the right knowledge means cohesion at scale, across categories, and across time and space.

Account Executive Mitch Pursell at Upland Software  put it well: 

 “Everyone gets excited about AI. The algorithms. The automation. The promise of faster, smarter decisions.

But here’s the part that often gets overlooked: if your knowledge is outdated, your AI will fail — quietly, but completely.”

Let me end with an example that demonstrates the scope of knowledge that’s required for true understanding. Archaeologists and geneticists are able to shed more light on the peoples of the ancient world, including the Etrurians in what is now Italy and the Carthaginians, who had a presence on some of the islands of the Mediterranean, not to mention many other peoples of whom evidence exists but is not plentiful.

To understand what was going on during the heyday of Ancient Greece and Rome, it’s not sufficient to read what survives of Greek and Roman writings. We have so much information to add now.  Today we have all the discoveries of archaeologists. And we have the ability to examine ancient DNA.

To gain an understanding of that history, all this currently available, relevant information needs to be out on the table, sequenced timewise and correlated relative to who was in what place, when, where and why. 

 Like Classicists in the era of genome sequencing, businesspeople need, not only “the data”, but the right connections associated with that data out on the table in front of them, put together in a way that makes the most sense, to understand better. This way, professors can engage and inform their students, and businesspeople can make the best decisions they can.

And that’s why agents need their own relevant subset of what the businesspeople have, to do the bidding of the business. That’s why cohesion is important. Cohesion is the belonging togetherness, the logical connectedness, that agents need to be successful, reliable and predictably safe.