Don’t Allow AI Agent Hype to Outpace Reality

Here’s a revised version of your content with the promotional calls-to-action and external website mentions removed or rephrased:

—

Let’s start with the term “agent” itself. Right now, it’s being applied to everything from simple scripts to sophisticated AI workflows. There’s no shared definition, which leaves ample space for companies to market basic automation as something much more advanced. This kind of “agentwashing” not only confuses customers but also leads to disappointment. While we may not need a rigid standard, we do require clearer expectations regarding what these systems are supposed to accomplish, how autonomously they operate, and how reliably they perform.

Reliability is the next significant challenge. Most of today’s agents are powered by large language models (LLMs), which generate probabilistic responses. These systems are powerful but also unpredictable. They can fabricate information, stray off course, or fail in subtle ways—especially when tasked with completing multistep processes that involve external tools and chaining LLM responses together. A recent incident illustrated this: Users of a certain AI programming assistant were incorrectly informed by an automated support agent that they couldn’t use the software on more than one device, leading to numerous complaints and subscription cancellations. However, the policy didn’t exist; it was a fabrication by the AI.

In enterprise settings, such errors can lead to substantial repercussions. We must shift away from treating LLMs as standalone products and begin developing comprehensive systems around them—systems that address uncertainty, monitor outputs, manage costs, and incorporate safety and accuracy measures. These practices will help ensure outputs align with user requirements, comply with company policies regarding information access, and respect privacy considerations.

Some organizations are already taking steps in this direction, creating structured frameworks that integrate language models more deliberately. Our latest initiative is designed for enterprise reliability, combining LLMs with company data, public information, and other tools to ensure consistent and dependable outputs.

However, even the most advanced agent won’t be effective in isolation. For the agent model to function optimally, different agents must collaborate—whether that’s booking travel, checking the weather, or submitting expense reports—without the need for constant human oversight. This is where a proposed universal protocol comes into play, intended to allow agents to communicate capabilities and delegate tasks effectively. In theory, it is a sound approach, but in practice, it still has limitations. While it specifies how agents interact, it doesn’t clarify the meaning behind their actions. For instance, if one agent mentions it can provide “wind conditions,” another must guess if that’s relevant for assessing weather on a flight route. Absent a shared vocabulary or context, coordination can become fragile. This issue has been encountered in distributed computing before, and solving it at scale remains a complex challenge.

—

Feel free to let me know if you need further adjustments!