Agentic AI is having a moment. Every product announcement now includes an "agent" somewhere in the slide deck. But beneath the marketing, there is a small canon of research papers that defined what agents actually do: reason, act, use tools, remember, plan, and collaborate. These five papers are the closest thing the field has to a textbook, and they are genuinely fun to read.
ReAct: Synergizing Reasoning and Acting
ReAct, from researchers at Princeton and Google, introduced the loop that now powers most agent frameworks. The model generates a thought, takes an action, observes the result, and repeats. This interleaving of reasoning and action beats chain-of-thought on tasks that require interaction with external systems, because the model can correct its plan based on real feedback rather than hallucinating a fixed sequence. If you understand ReAct, you understand the heartbeat of every modern agent.
Toolformer: Language Models Can Teach Themselves to Use Tools
Toolformer, from Meta, showed that a language model can learn when to call an external API, which API to call, what arguments to pass, and how to incorporate the result into its final answer. The training process is elegant: the model generates potential API calls, executes them, and keeps only the calls that reduce loss on the next token prediction. The paper covers calculators, search engines, translation systems, and calendars. It is the bridge between "LLM as text generator" and "LLM as system orchestrator."
Generative Agents: Interactive Simulacra of Human Behavior
This Stanford paper is the most entertaining on the list. The authors built a small society of AI agents inside a Sims-like environment. The agents wake up, make plans, remember past experiences, reflect on them, talk to each other, and coordinate future actions. The architecture combines memory streams, reflection, and planning into something that feels alive. If you want to understand why memory and continuity matter in agent design, read this paper.
Voyager: An Open-Ended Embodied Agent with Large Language Models
Voyager, from NVIDIA, is a lifelong learning agent that plays Minecraft. It writes its own code to solve tasks, stores successful programs in a skill library, and retrieves them for future use. Unlike agents that start from scratch every episode, Voyager accumulates competence. The skill library is the key insight: agents need a way to remember what they have already learned, not just what they have already seen.
AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation
AutoGen, from Microsoft Research, frames agentic AI as a conversation between multiple specialized agents. One agent writes code, another reviews it, a third executes it, and a fourth summarizes the result. The framework handles the routing, termination, and human handoff logic that turns a single model into a team. Most production agent systems today use some variant of this multi-agent pattern.
Why These Five
Together, these papers cover the full agent lifecycle: reasoning and action (ReAct), tool use (Toolformer), memory and social behavior (Generative Agents), lifelong learning (Voyager), and multi-agent collaboration (AutoGen). You do not need to memorize the implementation details. Understand the core ideas, and every agentic system you encounter will look like a recombination of these same pieces.