Agentic AI Is Neither Intelligent nor an Agent (gfrm.in)

A recent exploration into the term "agentic AI" reveals significant shortcomings in the current industry definition of AI agents. The author’s work involved creating agents, such as a Bayesian learner and an evolutionary system, to genuinely embody the characteristics of an agent—maintaining beliefs, setting goals, and making decisions based on those beliefs. In examining LangChain's ReAct agents, the author argues that these systems do not truly function as agents, as they lack a robust model of the world, decision-making capability, and the ability to learn from experiences. This critique emphasizes the difference between mere programming logic and true agency in AI. The implications are substantial for the AI/ML community. Through experiments comparing the traditional LangChain ReAct agent with a newly developed Bayesian agent named Credence, the author illustrates key differences: while the LangChain agent achieved higher accuracy in answers, it ultimately performed worse in point-based tasks due to its indiscriminate querying strategy. Credence, by contrast, methodically evaluated whether the value of information was worth the cost, demonstrating that a genuine agent should be able to adapt and respond to changing conditions without losing sight of its goals. This conversation urges the industry to rethink its approach to AI agents and raise the bar for what constitutes true agency in artificial intelligence.