AI Letters #24  ·  LLM Showdown #15

ReAct Agents: From Paper to Production

How the Reasoning + Acting paradigm moved from a 2022 Princeton paper to three competing framework implementations — and what each prioritized.

2022
ReAct paper published
(Yao et al., Princeton + Google)
frameworks benchmarked
SynapseKit · LangChain · LlamaIndex
6→19
lines of code range
for a working ReAct agent
Click any event to expand details
2022 — October Paper
Yao et al. — "ReAct: Synergizing Reasoning and Acting in Language Models"
Princeton + Google Brain publish the ReAct paper. Core insight: interleaving chain-of-thought reasoning with tool-use actions outperforms either alone on knowledge-intensive tasks. Tested on HotpotQA, Fever, ALFWorld. The Thought/Action/Observation loop becomes the canonical agent pattern.
2023 — Q1 Framework
LangChain ships create_react_agent + AgentExecutor
LangChain implements ReAct as a first-class abstraction. The @tool decorator makes any Python function an agent tool. AgentExecutor exposes verbose, return_intermediate_steps, handle_parsing_errors — full loop observability from day one. Becomes the de facto standard for production ReAct agents.
2023 — Q2 Framework
LlamaIndex adds ReActAgent with FunctionTool abstraction
LlamaIndex brings ReAct to its document-centric ecosystem. FunctionTool.from_defaults() wraps any Python function. The agent is designed to work seamlessly with existing QueryEngines and RetrieverTools — making RAG + agent pipelines composable. response.sources surfaces the full reasoning trace.
2024 Framework
Agentic frameworks proliferate — ReAct becomes table stakes
AutoGPT, CrewAI, AutoGen, LangGraph, LlamaIndex Workflows — every framework ships a ReAct variant. The differentiation shifts from "does it support agents?" to "how composable is the loop?", "what's the observability story?", and "how many built-in tools come included?"
2025 Framework
SynapseKit 1.x ships ReActAgent with 18 built-in tools
SynapseKit takes the batteries-included approach: CalculatorTool, DateTimeTool, DuckDuckGoSearchTool, WikipediaTool, PythonREPLTool, ArxivSearchTool, and 12 more — all importable without any tool-definition code. Trade-off: no verbose, no return_intermediate_steps. The loop runs opaque by default.
2026 — April Benchmark
LLM Showdown #15 — ReAct agent construction benchmark
Notebook #15 measures lines of code (imports + setup), built-in tool inventory, and loop control parameters across all three frameworks. SynapseKit: 6 lines, 18 built-in tools, 4/6 loop control. LangChain: 19 lines, 15 built-in tools, 6/6. LlamaIndex: 13 lines, 9 built-in tools, 6/6. The observability gap is the headline finding.
Framework Design Philosophy — What Each Optimized For
www.engineersofai.com  ·  AI Letters #24  ·  LLM Showdown #15