AI Agents Papers to Read in 2026: The Core Reading List

If you are past the courses and want to read the AI agents papers that the field is actually built on, this is the list. It is the short shelf of research that every framework, every course, and every engineering essay on this site keeps referring back to. Six papers, each one a fork in the road where agent design split off from plain prompt engineering. Treat it as the reading list that turns a vague sense of how agents work into the precise vocabulary the original authors used.

The list is organized by what each paper contributes, grouped into reasoning foundations and then the applied agent papers. It is the companion to the 12-post Learn AI Agents series: where the series teaches the how, these six papers are the why behind it. The seventh paper, ReAct, is the foundational one and gets its own full deep dive, so it is linked at the end rather than repeated here.

The reasoning foundations: AI agents papers that taught models to think before they act

Before an agent can call a tool or reflect on a mistake, it has to be able to reason through a problem step by step. Three of the six papers established that substrate. They are the AI agents papers you read first if you want to understand why chain of thought, search, and self critique show up inside every modern agent loop.

• Chain-of-Thought Prompting (Wei et al., arXiv). The paper that showed a model reasons better when it is asked to spell out its intermediate steps instead of jumping to an answer. Almost every later agent technique, ReAct, Reflexion, Tree of Thoughts, assumes this trace exists, which is why this one belongs at the top of the reading order. Read Chain-of-Thought Prompting on arXiv.
• Tree of Thoughts (Yao et al., arXiv). Takes the chain of thought idea and generalizes it into a search, where the model explores multiple branches, evaluates them, and backtracks when a path looks weak. It is the bridge between prompting and classic search, and a useful pattern to reach for before you wire in a heavier agentic planner. Read Tree of Thoughts on arXiv.
• Reflexion (Shinn et al., NeurIPS 2023). Adds a self-reflection loop in which the agent writes a short note about why its last attempt failed and feeds that note back into the next try. It is one of the cleanest demonstrations that verbal reinforcement, the model critiquing itself in natural language, can substitute for a gradient update. Read Reflexion from NeurIPS 2023.

The applied agent papers: tools, memory, and the knowledge layer

Once a model can reason, the next questions are how it reaches outside its own weights, how it behaves over time, and where its grounding comes from. These three papers answer those questions. They are the AI agents papers that moved the field from a model that thinks to a system that acts, remembers, and retrieves.

• Generative Agents (Park et al., arXiv). The small-town simulation that showed agents could carry persistent memory, reflect on their days, and produce believable social behavior over time. It is the paper most often cited when someone wants to argue that an agent needs a memory and reflection architecture, not just a prompt. Read Generative Agents on arXiv.
• Toolformer (Schick et al., NeurIPS 2023). Taught a model to decide for itself, mid generation, when to call an external API and how to fold the result back into its text. The entire premise of tool-using agents, that the model is the orchestrator and the tools are slots it fills in, comes from this work. Read Toolformer from NeurIPS 2023.
• Retrieval-Augmented Generation Survey (Gao et al., arXiv). The standard survey for the retrieval layer that most production agents sit on top of, covering chunking, embedding, indexing, and the retrieval generation split in one place. Read it last, once you understand the agent loop, because it explains the knowledge half that reasoning and tool use depend on. Read the RAG Survey on arXiv.

How these six fit with ReAct

Read together, the six sketch a complete picture. Chain-of-Thought Prompting and Tree of Thoughts give the model a way to reason and search. Toolformer gives it hands. Generative Agents and Reflexion give it memory and the habit of self critique. The Retrieval-Augmented Generation Survey gives it a knowledge base to draw from. The one missing piece is the loop that interleaves reasoning with acting, and that is ReAct, the foundational agent paper. ReAct is deep enough to deserve its own walkthrough, so it lives in a separate deep dive rather than a single bullet here. If you finish these six and want the paper that ties reasoning and action together, that is the next read.

If you only have time for two of these AI agents papers, start with Chain-of-Thought Prompting for the reasoning substrate and Toolformer for the tool use premise. Together they cover the two moves, think then call, that every later agent paper composes. Tree of Thoughts and Reflexion extend the first move; Generative Agents and the RAG Survey extend the second. Read them in roughly that order, then finish with the ReAct deep dive to see the loop closed. For the long tail beyond these six, the LLM Agent Paper List indexes hundreds more as the companion to an 86-page survey paper. It has not been refreshed since September 2025, so treat it as an archive and verify anything you plan to cite.