ReActAgent

ReActAgent implements the Reasoning + Acting pattern. It works with any BaseLLM — no native function calling required.

How it works

Each iteration:

1. LLM generates a Thought, Action, and Action Input
2. The named tool is called with the input
3. The Observation (tool result) is appended to the scratchpad
4. Repeat until the LLM produces a Final Answer

Thought: I need to calculate 2 ** 10.
Action: calculator
Action Input: 2 ** 10
Observation: 1024

Thought: I now know the final answer.
Final Answer: 2 to the power of 10 is 1024.

Usage

from synapsekit import ReActAgent, CalculatorTool, WebSearchTool
from synapsekit.llm.openai import OpenAILLM
from synapsekit.llm.base import LLMConfig

llm = OpenAILLM(LLMConfig(model="gpt-4o-mini", api_key="sk-..."))

agent = ReActAgent(
    llm=llm,
    tools=[CalculatorTool(), WebSearchTool()],
    max_iterations=10,
)

answer = await agent.run("What is the population of France divided by 2?")

Streaming

async for token in agent.stream("What is 99 * 99?"):
    print(token, end="", flush=True)

Intermediate tool calls run silently; the final answer is streamed.

Inspecting steps

await agent.run("compute something")

for step in agent.memory.steps:
    print(f"Action: {step.action}")
    print(f"Input:  {step.action_input}")
    print(f"Result: {step.observation}")

Parameters

Parameter	Type	Default	Description
llm	BaseLLM	required	Any LLM provider
tools	list[BaseTool]	required	Available tools
max_iterations	int	10	Max Thought→Action cycles
memory	AgentMemory | None	auto	Custom memory instance