A Brief History of Agentic Loops: From ReAct to Claude Code

Milestone Summary

Date	Milestone	Key Contribution
2022.10	ReAct paper	Formalizes the Thought-Act-Observe pattern
2023.03	AutoGPT / BabyAGI	Exposes infinite-loop and hallucination-cascade failure modes
2023.06	Lilian Weng’s post	Planning · Memory · Tool Use three-axis framework
2024.12	Anthropic BEA guide	Production-grade workflow and agent design principles
2025.07	Ralph technique	Fresh-context + single-purpose loop principle
2025~	“Loop Engineering”	Loop design recognized as an independent engineering discipline

Why the History Is Short

The history of agentic loops is surprisingly compressed. The ReAct paper that formalized the core idea appeared in October 2022. Practical engineering principles were consolidated in fewer than three years. Packed into that span are: academic research, the spectacular failure of viral open-source projects, the maturation of production systems, and the emergence of a new engineering paradigm.

Agentic Loop Timeline
──────────────────────────────────────────────────────────────
2022.10  ReAct paper — Thought · Act · Observe pattern
         ALFWorld +34%, WebShop +10% vs. RL baselines

2023.03  AutoGPT / BabyAGI go viral
         Infinite loops and hallucination cascades exposed

2023.06  Lilian Weng "LLM Powered Autonomous Agents"
         Planning / Memory / Tool Use three-axis framework

2024.12  Anthropic "Building Effective AI Agents"
         Workflow vs. agent distinction, "start simple" principle

2025.07  Geoffrey Huntley's "Ralph" technique
         while :; do cat PROMPT.md | claude-code; done

2025 ~   "Loop Engineering" term emerges (Osmani, Cherny)
         Loop design recognized as an independent discipline
──────────────────────────────────────────────────────────────

2022.10 — ReAct: Combining Reasoning and Acting

Shunyu Yao et al.’s ReAct (arXiv:2210.03629) was the first paper to systematize the pattern of interleaving Reasoning and Acting. The core idea: let the model cycle through Thought → Act → Observe. Compared to reasoning alone or acting alone, this interleaving produced significantly better results.

The empirical evidence was compelling. On ALFWorld, ReAct beat the reinforcement learning baseline by +34 percentage points absolute. On WebShop it gained +10%. Another major advantage was transparency: because the model articulates intermediate thoughts in natural language, the reasoning trace is human-readable — making it far easier to diagnose failures. ReAct became the template for almost every agentic framework that followed.

2023.03 — AutoGPT and BabyAGI: Excitement and Failure

In March 2023, AutoGPT and BabyAGI were released as open-source projects and attracted millions of users almost overnight. The concept was revolutionary on the surface: GPT-4 autonomously decomposing goals, searching the internet, executing code, and iterating.

The excitement did not last. In practice, loops frequently ran indefinitely without stopping. Early hallucinations propagated into subsequent steps, compounding into completely wrong trajectories. Costs exploded. Two failure modes became impossible to ignore: the infinite loop and the hallucination cascade. These failures became the negative lesson that taught the community why loop control and termination conditions matter as much as the loop itself.

2023.06 — Lilian Weng’s Framework

OpenAI’s Lilian Weng published “LLM Powered Autonomous Agents” in June 2023, systematically organizing everything known at that point. She classified agentic systems along three axes — Planning, Memory, and Tool Use — and this taxonomy became the reference vocabulary for subsequent discussion.

Particularly influential was the categorization of planning approaches (ReAct, Reflexion, Chain-of-Thought, Tree of Thoughts) under the planning axis, and the distinction between short-term memory (context window) and long-term memory (external stores) under the memory axis. This post standardized the language of agentic system design.

2024.12 — Anthropic’s Production Principles

In December 2024, Anthropic published “Building Effective AI Agents” — not a research paper or conceptual survey, but a guide to building production systems. It articulated the workflow-versus-agent distinction, the “start simple” recommendation, and five foundational workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer).

The guide moved the conversation from “what agents can do” to “how to build agents properly.” Its insistence on not over-autonomizing — give agents appropriate constraints, not unlimited freedom — directly reflected the lessons of the AutoGPT failures.

2025.07 — Ralph: The Simplest Possible Loop

Geoffrey Huntley’s Ralph technique reduced the agentic loop to its most minimal form:

# The Ralph core loop — the simplicity is the point
while :; do cat PROMPT.md | claude-code; done

This looks almost trivially simple, but it encodes several important principles. Context is reset on every iteration (fresh context). A stable spec document (PROMPT.md) is re-injected each time. Each loop run pursues a single goal. The hallucination-accumulation problem is solved by discarding context rather than preserving it. “Run the loop — but keep the loop simple” was the counterintuitive wisdom.

2025 Onward — “Loop Engineering” Emerges

From 2025, Addy Osmani and Boris Cherny (lead of Claude Code) articulated a perspective that has been spreading: loop design itself is an independent engineering discipline. Moving beyond “writing a good prompt” or “structuring a good context window,” the focus shifted to the structure of the loop itself — its termination conditions, state management, cost controls, and reliability — as the primary engineering competency.

What this compressed history teaches us is simple: agentic loops began in theory, were tempered by failure, extracted principles from that failure, and are now in the process of systematizing those principles. The next section explores those principles one anatomical layer at a time.

Loading quiz…