Why are AI agents useful now?

Prerequisites

In the late 2025, three capabilities of LLM-based AI agents have surged, making them far more attractive for a growing range of applications:

• the ability to take actions via tool calling
• the ability to "plan" a sequence of actions
• the ability to process relatively large amounts of unstructured multimodal data

Even though agents still make mistakes, each of these aspects has improved dramatically.

Enterprise perspective

From an enterprise perspective, this means that many routine intelligence jobs can now be automated to a large degree. The key question is how much risk the stakeholders are willing to accept. Risk comes mainly from two sources:

• hallucinations / confabulations (see prerequisites for reading),
• failing to follow the instructions (an aspect of misalignment).

Arguably, the safest AI agent use cases usually share the following traits:

1. Factual truthfulness is nice but not essential; it’s not catastrophic if some (or even all) facts are wrong.
2. The benefits of getting something useful outweigh by a considerable margin the harm of getting wrong or misleading information.
3. Generated outputs should not be the sole source of decision‑making and should not be part of a fully automated pipeline.
4. If factual truthfulness is required, expert knowledge to assess the agent’s output must be readily available and actively used.

Depending on your risk tolerance, you can relax these rules of thumb. When done right, AI agents can boost consistency, cut onboarding time, and deliver cost savings.

Societal perspective

From a societal perspective, AI applications fall into three broad categories: the good, the bad, and the ugly.

The good side includes

• personalized medicine 🩺
• tailored education 🎓
• robotics for hazardous jobs 🤖🛠️

There is more, but these three have arguably the most potential.

The bad and the ugly sides have too many nuances to discuss and merit their own intro series. Stay tuned for that.