The State of AI Agents
Cutting edge papers with shocking techniques featuring even more shocking results are my favorite thing to breakdown.
Lots of those lately in AI / ML.
I just spent the weekend reading an absolutely incredible (and massive, 264 pages) paper that surveyed the current state of AI Agents. It was packed with insights I hadn’t read anywhere else yet. It also made me even more excited about our agentic renaissance.
To get started I used NotebookLM from Google to create an audio overview of the paper, just to get my brain warmed up. Here it is:
The paper was too big for Gemini to read it entirely. Context windows are still a thing in 2025, even with the best model in the game. Let’s unpack the paper and get into the most exciting findings.
The Rapidly Rising Power of AI Agents
The paper makes the case that while LLMs represent a significant leap in AI, demonstrating capabilities in language, reasoning, and perception, they are merely foundational components—like engines—for truly intelligent agents, which are analogous to complete vehicles. The paper proposes a modular, brain-inspired architecture for designing these agents, integrating principles from cognitive science and neuroscience.
The survey systematically explores three key areas:
Core Components: It maps the cognitive (learning, reasoning, planning), memory (short-term, long-term, sensory), perceptual, operational (action), world modeling, reward processing, and even emotion-like modules of AI agents onto analogous human brain functions. It delves into how these components function and interact within the agent framework.
Self-Evolution: It discusses mechanisms for agents to autonomously enhance their capabilities and adapt over time. This includes continual learning, automated optimization strategies like AutoML, and using LLMs themselves as optimizers for prompts, workflows, or tools. Both online and offline self-improvement methods are considered.
Multi-Agent Systems: The paper examines collaborative and evolutionary aspects of systems with multiple agents. It investigates collective intelligence, cooperation, competition, communication protocols, and societal structures emerging from agent interactions, drawing parallels to human social dynamics.
Ultimately, the paper aims to synthesize insights across disciplines, identify research gaps, and encourage innovations that advance AI agent capabilities while ensuring they align with societal benefit.
While the concept of agents isn't new, this paper proposes a comprehensive, modular architecture explicitly inspired by human brain functions. It systematically maps agent components like memory (distinguishing between sensory, short-term, long-term), cognition (learning, reasoning, planning), world models, reward systems, emotion modeling, perception, and action to their cognitive or neurological counterparts. This structured, holistic view contrasts with earlier agent designs that might have focused on optimizing a single function (e.g., pathfinding) or lacked this explicit cognitive parallel and modular integration.
We see several breakthroughs and evolutions explained inside this paper.