Agent Frameworks

By mid-2023, the conceptual recipe for LLM agents — ReAct-style Thought/Action/Observation loops, tool-use APIs, retrieval, memory — was well understood. What followed was an explosion of agent frameworks that orchestrated these primitives into reusable building blocks: AutoGPT, BabyAGI, LangChain Agents, LangGraph, CrewAI, AutoGen, OpenAI Swarm. Each took a different bet on what the right abstraction is. Most are still finding their footing as of 2025; the space is unsettled.

AutoGPT and the autonomous-agent moment (March 2023)

AutoGPT (Significant Gravitas, March 2023) was the first viral autonomous agent. Give it a goal in natural language; it would loop indefinitely, planning sub-goals, calling tools (search, file read/write, code execution), updating memory, and self-evaluating progress. BabyAGI (Yohei Nakajima, April 2023) was a similar but simpler system focused on a task-list-driven loop.

Both went viral on Twitter/X in spring 2023. The pattern — give the LLM a goal and let it run — captured imaginations. The reality was that AutoGPT-style autonomous agents:

• Got stuck in loops on most non-trivial tasks.
• Hallucinated tool calls at high rates.
• Burned tokens without making progress.
• Failed silently without robust error recovery.

By summer 2023 the autonomous-agent hype had cooled. The lessons survived: GPT-3.5 / GPT-4 weren't capable enough to run open-ended autonomous loops yet, but the recipe was right. Capability eventually caught up.

LangChain — the framework that ate the early ecosystem

LangChain (Harrison Chase, Oct 2022) became the dominant agent + RAG framework of 2023, with bindings for every LLM provider, vector store, retriever, and tool. Its core abstractions:

• Chains — declarative pipelines of LLM calls and transformations.
• Agents — ReAct-style loops with tool registries.
• Memory — conversation buffers, summary memory, retrieval-augmented memory.
• Retrievers — abstraction over vector stores for RAG.

LangChain's reach was unprecedented: by mid-2024, ~70% of LLM-app blog posts used it. The criticism was equally widespread — too many abstractions, leaky generalisations, hard-to-debug stack traces. LangChain reorganised in 2024 into a smaller core + LangGraph + LangSmith + LangServe ecosystem, partly addressing these critiques.

LangGraph (2024) is the framework's modern recommendation: explicit state graphs (nodes are functions, edges are routing decisions) instead of declarative chains. The state-machine view turned out to be a much cleaner abstraction for agent loops than chained transformations.

LlamaIndex — RAG-first

LlamaIndex (Jerry Liu, late 2022) competed with LangChain for the same audience but emphasised structured RAG — sophisticated indexing, retrieval strategies, and document handling. By 2024 the line had blurred; LlamaIndex added agent capabilities, LangChain added structured RAG.

AutoGen and CrewAI — multi-agent orchestration

A second wave of frameworks targeted multi-agent patterns:

• AutoGen (Microsoft, 2023) — a framework for orchestrating multiple LLM agents that converse with each other. "Pair programming" agent conversations, debate-based reasoning, planner-executor splits.
• CrewAI (2023) — agents as "crew members" with distinct roles, collaborating on tasks.
• OpenAI Swarm (Oct 2024) — a minimal example library for multi-agent orchestration; explicitly not a production framework.

The multi-agent thesis is that complex tasks benefit from specialisation — a researcher agent, a writer agent, a critic agent, etc. Empirically, the gains are uneven; sometimes a single capable model with the right prompt does as well or better. Multi-agent systems work well when:

• Different agents have access to different tools or contexts.
• Adversarial structure is useful (critic vs writer).
• Parallelism is important.

They fail when the orchestration cost (extra LLM calls, coordination prompts) exceeds the specialisation benefit.

Production agents in 2024-2025

The frontier-model providers built agentic capabilities directly into their offerings:

• OpenAI Assistants API (Nov 2023) — managed thread state, tool registry, file search, code interpreter as built-ins.
• Anthropic Computer Use (Oct 2024) — Claude with browser/computer control built into the API. See computer use.
• OpenAI Operator (Jan 2025) — autonomous browser agent.
• Claude Code (Feb 2025) — CLI-based coding agent.
• Gemini live agentic capabilities — search, action.

The pattern is increasingly: provider gives you a strong agentic LLM with tool-use and computer-use baked in; the developer's job is plumbing tools, defining policies, and writing system prompts. The "framework" is becoming thinner as models become more capable.

What agent frameworks are for

The right framework choice in 2025 depends on use case:

• Simple ReAct loop — just write Python with the provider's tool-use API. Frameworks add cost without value.
• Complex routing / state graph — LangGraph or similar.
• Multi-agent collaboration — AutoGen, CrewAI for prototyping; rarely production.
• Document QA / RAG — LlamaIndex or LangChain's retrieval components.
• Production deployment — provider-native (Assistants, Bedrock Agents) for stability.

The framework market is consolidating. By 2026, expect a smaller set of established frameworks plus provider-native agents.

What to read next

• Tool Use — the underlying primitive.
• Coding Agents — the most-developed agent application.
• Computer Use — the next-generation agentic action.