Taking a brief detour into economics, an agent is any person or organization that is given agency, i.e., some combination of freedom and responsibility to represent another person or organization. Typically, an agent is also given goals that direct their behavior and provide a basis for evaluating their performance. In computer science, we extend this concept to intelligent agents or AI agents, which are software entities that act on behalf of another person or organization. We often imply that AI agents are autonomous, meaning that they are able to independently observe the world around them, reason about it, and act upon it — then repeating that sequence in what we call the observe-reason-act loop. This is also called the sense-think-act loop. If a human is involved in any of those phases, we refer to that as an AI agent with human-in-the-loop. A further enhancement to AI agents includes a fourth phase, learning. If the AI agent is able to learn from the effects of its actions and therefore improve its reasoning performance, it is called a learning AI agent.

With a large language model, we can specify goals to be achieved by an AI agent, which then uses an AI planning algorithm to generate a sequence of tasks based on the actions available to it through the language model. The agent then starts executing the tasks and iteratively evaluates the output of those tasks to reason about whether the task resulted in progress towards the goals as expected. If yes, it continues down the planned execution sequence; if not, it replans to regenerate the remaining task sequence. This process leads the agent to efficiently and robustly achieve the specified goals.

First, generative AI agents are able to perform tasks by generating thoughts that guide their actions and by producing content in the form of text, images, etc. through the use of LLMs. Then, interactive AI agents further extend this generative capability to more effectively use other agents in their environment. The other agents may be several human experts or other AI agents. Thus, an interactive AI agent must maintain awareness of other agents in its environment, discover their specialized capabilities that could complement its own, and work with them to better accomplish its own tasks.

Artian’s approach to AI agents emphasizes our view that not all knowledge in the world will be readily available to public large language models like GPT-x. Private and premium knowledge will be made accessible to AI models at significant costs. Whether an AI agent is willing to pay that cost depends on the perceived value of that knowledge to the goals that it is trying to achieve. Artian’s self-learning AI agents are able to make this determination autonomously through analysis of ongoing task execution, thus resulting in significant cost advantages in knowledge acquisition. We can also inject human supervision into this process at various stages, considering the deployment scenario, to ensure smooth adoption.

Frameworks for autonomous agents and multi-agent systems have existed in AI literature and research software for over two decades. However, there is currently no standard enterprise-grade platform for AI agents. Different model ecosystems and open source frameworks for agentic AI are evolving concurrently. We expect that several frameworks and platforms will co-exist, and work together through protocols like MCP and A2A for the foreseeable future.

There are many emerging frameworks and platforms that promise an agentic reinvention of your applications. However, all of these are either almost entirely LLM-driven and inherently unreliable for mission-critical autonomous use, or they require explicit and complex programming that is out-of-reach for business developers. Artian uniquely blends the benefits of LLMs for productivity and the structure of workflows for reliability.

That depends on your preferences and your specific business goals. While we strongly believe that autonomous agents can eventually accomplish many business tasks independently, we also believe that AI should be introduced into our workplaces gently. Active supervision by human experts is often critical to the success of agents and also to the success of the people involved towards their business and career goals.