AI 2025 - The Year of the Agent

The path to AGI, diversity of subsystems, and complexity of coordination.

Where are we?

"A lot of what we are focused on is giving every creator, and every small business the ability to create agents for themselves." Founder Chairman and CEO of Meta, Mark Zuckerberg.

Tech leaders agree. 2025 is the year of the AI agent. Agents in the virtual worlds, Salesforce agents conducting business, agents embedded in social media, sophisticated chatbots, and more. The progression of this type of functioning will also make its way into robots where versions likely already exist.

"All of a sudden, as a CEO, I'm not just managing human beings, but I'm also managing agents" – it's not a fantasy of the future, "it is what is happening right now." Salesforce's Co-Founder and CEO, Marc Benioff

It's all already underway, and the sophistication will increase rapidly.

"I think the models are definitely going to get better at reasoning, completing a sequence of actions more reliably — more agentic if you will." Google CEO, Sundar Pichai

Suppose you think about how you interact with AI today. AI reacts to your prompting. AI agents are different in that they will have a task or job they are designed for, and they do it. Responding to events or changing conditions within the context they exist in.

Image created with Adobe Firefly and Photoshop

How did we get here?

Video games are half the reason I followed a career path in computers. In 2004, I was in my last year of tech school, and I chose to write a cribbage game for my final Visual Basic project. It was a game I had just learned, and my teacher said I couldn't do it. As a contrarian, that sealed the deal.

The challenge worked. I finished a functional game of cribbage with a UI that was on par with the version of Solitaire that came with Windows in those days. Representing a deck and shuffling was a parlor trick I had learned in high school. Simple logic and determination got me through counting points during gameplay. The chasm I did not have time to bridge was getting the PC opponent to choose or play its cards wisely.

I headed to Google to research. Soon, I was down the rabbit hole of artificial intelligence. Game development has always pushed the boundaries of technology. I found plenty of literature on NPC activity in the gaming field.

Outside of gaming, I spoke with a teacher who had developed a neural network in Java. I read about their use in image recognition, translation, and more. Some less technical articles about the subject spoke of the future of AI and how we might advance the field by looking to neuroscience for ideas.

If we use the neuroscience lens, most AI today uses neural networks analogous to human neurons networked together to do the things a brain does. There are key differences between how human brains function and neural networks in computing. There are also many similarities.

Where we go

We've advanced to a mixture of LLMs, some specialized to a task collaborating with more generalized LLMs. These are fed by other networks that translate inputs into a form that the LLMs can work with. Multi-modal AI means the inputs to a system can be visual, audio, text, or measurements from various sources. The outputs of this could be text, images, videos, sound, and control of software or hardware. These inputs and outputs are their own AI subsystem based on similar techniques.

Neuroscience and neural networks both are bottom-up views. As we enter the agent phase of AGI development, we should see a shift to top-down approaches. With human beings, the corollary is psychology. Psychology and neuroscience have progressed to understand human functioning from different angles. Psychology focuses on the higher layers of function organizing and driving surface behaviors. Neuroscience, at the most granular level, is about building subsystems from the ground up with building blocks of limited function. The two fields are largely independent, but we are increasingly starting to see some meeting in the middle.

Differences between human brains and neural networks are in complexity, scale, learning, context, and reasoning. Complexity and scale can be partially addressed by scaling of hardware. Challenges exist in getting both quantity and quality of data. Improvements in algorithms also have a role to play in processing the data more efficiently and creating a better model. This is the human neuroscience approach.

The rest of the differences between human intelligence and AI will be addressed by a completely different approach that is more aligned with the psychological model. Frameworks that organize AI subsystems bring capabilities for learning, context, reasoning, and functioning in dynamic situations over time.

Progress towards AGI will not be singular complex agents but rather systems of agents, some of which specialize in the orchestration of the agents. The Magentic One framework, built on the AutoGen framework, takes this approach. Platforms like LangGraph allow agents to time travel and memory to effectively process and use past information. It also has features designed around incorporating humans into the loop.

The good news is there are a lot of concepts to draw from in advancing these frameworks. Hardware resources and the quantity and quality of data gated the maturity of neural networks over the decades. Neither of these will be a problem in the advancements of agentic platforms. Progress for the short term, at least, is primarily limited by our creativity and willingness to try new things.

Breaking the barrier

These advances are both good and inevitable. The interesting angle is how the fear of out-of-control AI is slowly eroding. When ChatGPT first wowed the world, there was a lot of focus on strict walls of separation between our AI and the "control planes" of the world. A world of AI agents requires AI to have more control and autonomy.

This reality will be a focus point as we open up agent capabilities. The need for oversight, control, and management will come to the forefront. Current black box problems of understanding how or why the agent did what it did will need to be addressed.

Computers were said to be coming for our jobs, and they did. Now we do higher level jobs on computers. I don't know if we work for them or they work for us. The latest iteration is AI will take your job, which may be, but someone will need to manage the bot farms. It's time to dive in and skill up.