Author(s): Anubhav

Originally published on Towards AI.

ReAct, Toolformer, Reflexion, Voyager, and the framework paper almost nobody opens. You already ship all five, you just never read where they came from.

Nearly every AI agent shipping today traces back to five research papers, and all five came out in the same 11 months. Between October 2022 and September 2023 the field published the reasoning loop, the tool calls, the self-correction, the skill library, and the architecture that ties them together. Then it spent three years turning them into function names you call without reading.

After introducing the idea of a five-paper lineage, the article walks through each work and what survived into modern agent tooling: ReAct establishes the “think, act, observe, repeat” loop and later replaced brittle string parsing with structured function/tool calls; Toolformer demonstrates learning when to use tools (but is largely obsolete in training and single-step form); Reflexion adds retry-with-verbal critique so agents can recover from failures, foreshadowing judge/verification pipelines while highlighting how reflection can also trap models in repeated dead ends; Voyager introduces a durable skill library so capabilities persist across tasks (while its autonomous curriculum mostly didn’t reach production); and CoALA provides the missing blueprint by organizing memory types and a unified decision loop, explaining why frameworks feel like siblings even when they don’t cite the paper. The piece concludes that recognizing these failure modes and components makes debugging and design more diagnostic—naming which paper’s limitation is showing up when an agent loops, miscalls tools, keeps repeating mistakes, or fails to retrieve the right capability.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI