AI Letters #24 · LLM Showdown #15 — Evidence Dashboard

ReAct Agents: Full Benchmark Results

Notebook #15 — Lines of code, built-in tool inventory, loop control, and custom tool cost across SynapseKit 1.4, LangChain 1.2, and LlamaIndex Core 0.14.

Lines of Code — Imports + Setup to a Working Agent

SynapseKit

3 imports + 3 functional

LlamaIndex

3 imports + 10 functional

LangChain

5 imports + 14 functional

* SynapseKit's 6-line count relies on built-in CalculatorTool and DateTimeTool. Task required two tools; SK ships both. LC and LI require tool-definition code.

Lines of Code (stacked)

Imports vs functional lines

Built-in Tool Inventory

Tools available without writing tool code

Loop Control Parameters

What the framework exposes to the caller for controlling and observing the ReAct loop

Parameter	SynapseKit	LangChain	LlamaIndex
max_iterations	Yes	Yes	Yes
early stop condition	Yes	Yes	Yes
handle_parsing_error	Yes	Yes	Yes
verbose (loop logging)	No	Yes	Yes
return_intermediate_steps	No	Yes	Yes
async support	Yes	Yes	Yes
Score (out of 6)	4 / 6	6 / 6	6 / 6

Custom Tool Definition Cost

Lines to define one new tool (when built-ins don't cover your use case)

Loop Control Score

Parameters exposed for observing and controlling the ReAct loop

Key Findings

Headline Number

SynapseKit's 6-line agent is real — but it only holds when your use case fits its 18 built-in tools. The moment you write a custom tool, LangChain and LlamaIndex cost the same or less.

Production Observability

LangChain and LlamaIndex both score 6/6 on loop control. SynapseKit's missing verbose and return_intermediate_steps are a meaningful liability when you need to debug a misbehaving agent in production.

The Quiet Winner

LlamaIndex at 13 lines with 6/6 loop control is the balanced choice — especially if you're already in the LlamaIndex ecosystem for RAG. Lower built-in tool count (9) is the only real gap.