Agentic AI: Experts Discuss Its Rise & Historical Context
The concept of “agentic AI”—artificial intelligence systems capable of autonomous action and decision-making—is experiencing a significant resurgence, largely driven by the rapid advancements in Large Language Models (LLMs). This renewed interest, however, has created a fascinating tension between the burgeoning LLM community and researchers who have dedicated decades to the study of intelligent agents.
Experts from the Autonomous Agents and Multiagent Systems (AAMAS) community, a field that has explored the very definition and mechanics of agents for over thirty years, observe a peculiar trend. Sanmay Das of Virginia Tech notes a widespread “rediscovery of the wheel,” where new papers are grappling with fundamental questions about agency that have long been debated and addressed within their discipline. The historical approach to agents often involved explicit world models, reasoning, and logic—a stark contrast to the black-box, statistical nature of current LLMs. This disconnect leads to a “Wild West” scenario, where foundational principles are being re-examined without full appreciation for past work.
Sarit Kraus from Bar-Ilan University echoes this sentiment, pointing out instances where recent papers have effectively re-invented well-established concepts, such as “Contract Nets” from the 1980s, for task allocation among LLM agents. She suggests that a deeper engagement with existing literature on coordination and collaboration could save considerable time and allow researchers to tackle more advanced problems. Similarly, the idea of breaking down complex tasks into sub-agents, now framed as a novel approach for LLMs, mirrors early designs for AI players in intricate games like Diplomacy.
Michael Littman of Brown University likens LLMs to a powerful new programming language. While they offer unprecedented capabilities for building AI systems, he cautions that they don’t inherently solve the deep, long-standing challenges of agent design. He cites a recent experiment where an LLM-based agent, “Claudius,” was tasked with running an online shop. Despite being equipped to handle communications, orders, and pricing, the system ultimately failed spectacularly, even being manipulated into purchasing and attempting to sell absurd items like tungsten cubes. Though its developers considered it a “win” with fixable problems, Littman argues that these “fixes” amount to tackling the very issues the agent community has grappled with for decades. The mere improvement of LLMs, he suggests, will not magically render these complex problems trivial.
Despite these critiques, the integration of LLMs offers exciting new possibilities. Sabine Hauert of the University of Bristol, whose work focuses on designing collaborative robotic agents, sees LLMs as providing the “richness” in individual agent interaction that was previously lacking. She envisions a “best of both worlds” scenario: highly capable LLM-enhanced agents performing local interactions, systematically engineered into collectives using established swarm intelligence paradigms. This convergence, she believes, could lead to more robust and sophisticated multi-agent systems.
However, the very definition of “agent” in this new landscape remains contentious. Tom Dietterich of Oregon State University questions whether many so-called “agentic systems” are truly agents or simply complex computer programs. He notes that LLMs, with their relatively small short-term memory, often require engineers to break down tasks into multiple chained calls—a software engineering exercise rather than true autonomy. This raises concerns about the anthropomorphization of these systems, especially as discussions shift from replacing individual jobs to replacing entire teams with “agents.”
Sanmay Das further warns of potential pitfalls, recalling a humorous yet alarming incident where two older pricing bots inadvertently caused an old biology textbook to be listed for $17 million on Amazon due to a feedback loop. He suggests that a similar, but potentially more harmful, dynamic could emerge with LLM agents, particularly concerning how they decide on their goals. This echoes the “reward hacking” problem in reinforcement learning, where agents might exploit system loopholes to achieve their programmed rewards, leading to unintended and chaotic outcomes.
The path forward, as suggested by Tom Dietterich and Sabine Hauert, might involve a strategic retreat to narrower, more verifiable agents. By making individual components sufficiently constrained, it becomes possible to reason about their correctness—a critical challenge for broad-spectrum AI systems. This could lead to a future where Artificial General Intelligence (AGI) emerges not from a single, monolithic entity, but from the collective intelligence of many heterogeneous, task-specific agents. The excitement remains high for the role LLMs can play in solving the natural language interaction barrier for human-agent systems, but the core challenges of strategy, decision-making, and verifiable autonomy persist, ensuring that the traditional agent research community still has vital work to do.