Redefining Agents in the World of Gen AI

What is a “Real Agent”?
When I talk about a “real agent,” I’m referring to something much more sophisticated than your average chatbot. A real agent should have a Large Language Model (LLM)-driven persona built around specific roles and goals. This setup goes beyond simple Retrieval-Augmented Generation (RAG) by incorporating multiple tools and, crucially, some level of reasoning based on the persona.

A real agent isn’t just about responding to queries; it’s about understanding context, making informed decisions, and executing tasks with a certain degree of autonomy. It’s this combination of persona, tools, and reasoning that, in my view, defines a real agent.

The Multi-Agent Conundrum
The term “multi-agent” is another area where definitions can get murky. For some, a system becomes multi-agent as soon as it includes more than one agent. However, I believe a true multi-agent system requires more than just multiple agents existing side by side. They need to communicate and collaborate towards a common goal, following a sequence of interactions.

In my perspective, every agent is, in essence, a multi-agent. Take an HR agent, for example; it should consist of 10-20 different internal agents interacting in a coordinated manner to achieve specific HR-related objectives. This creates a multi-agent system within a single agent. Then, this HR agent interacts with other agents (which are also multi-agent systems), forming a larger, more complex multi-agent ecosystem.

The Challenge of Reasoning
One of the biggest challenges in developing both single and multi-agent systems is reasoning, or the creation of an action plan. While there is considerable research focused on improving reasoning within individual agents, the larger issue lies in the collaboration and negotiation between agents within a multi-agent system.

Effective multi-agent systems require agents to work together seamlessly, negotiating and collaborating to achieve higher-level goals. This aspect of agent design is still a significant hurdle, but it’s one that holds the key to realizing the full potential of multi-agent systems.

Beyond Frameworks
There are numerous frameworks available today for creating both agents and multi-agent systems, often serving as a layer of packaging on top of existing generative AI tools. While these frameworks are useful, they are still relatively simplistic. My goal is to push beyond these limitations and offer a new level of agentic experience to our clients, in a pragmatic way.

I envision a future where agent-driven applications are the norm. However, defining these agents precisely and understanding their interactions within multi-agent systems is crucial. Perhaps we should start calling an LLM persona-based, action-oriented solution simply an “Agent.” This agent, composed of multiple internal agents, would then form a multi-agent system, capable of interacting with other agents at an external level.

In my experience, a good way of starting is to find that tangible use case which gives business value from the very start. Some tend to start with a PoC to prove a concept, but I’d recommend having the PoV mindset – Proof of Value. And then just do it! AI and Agents is something we all need to act on, and planning building roadmaps will prevent us from moving forward and learning. That’s the way I truly recommend; start by doing, focus on business value, go all the way to production from the start, even if with a very small user group, and then integrate your strategy and architecture while doing – and deliver business value.

The Future of Agentic Experiences
Ultimately, we might just see this new agentic experience as the new normal, with all these sophisticated solutions eventually being referred to as intelligent apps or just apps, as we’ve always done. But for now, it’s essential to continue refining our definitions and understanding of what makes a true agent and how multi-agent systems should function.

Feel free to challenge these ideas. The future of agents in generative AI is remarkable and still wide open, and it’s up to us jointly to shape it with new understandings and innovative approaches.