Reasoning Extract Agents

Chain a reasoning model with a smaller agent to extract concise answers

A reasoning agent breaks problems into steps; a follow-up agent extracts a concise answer from that chain.

from praisonaiagents import Agent, Task, AgentTeam, OutputConfig

reasoner = Agent(
    name="Reasoner",
    instructions="Think step by step.",
    llm="deepseek/deepseek-reasoner",
    output=OutputConfig(reasoning_steps=True),
)

extractor = Agent(
    name="Extractor",
    instructions="Answer briefly using the prior reasoning.",
    llm="gpt-4o-mini",
)

team = AgentTeam(
    agents=[reasoner, extractor],
    tasks=[
        Task(description="How many r's in 'Strawberry'?", agent=reasoner),
        Task(description="From the reasoning, how many r's?", agent=extractor),
    ],
)

team.start()

The user asks a hard question; a reasoner thinks step-by-step and an extractor returns a concise answer.

Quick Start

1. Simple Usage

   Enable reasoning output on a reasoning-capable model:

   from praisonaiagents import Agent, OutputConfig
   
   agent = Agent(
       name="Reasoner",
       instructions="Think step by step before answering.",
       llm="deepseek/deepseek-reasoner",
       output=OutputConfig(reasoning_steps=True),
   )
   
   agent.start("How many r's in 'Strawberry'?")

2. With Configuration

   Chain two agents — reasoner then extractor:

   from praisonaiagents import Agent, Task, AgentTeam, OutputConfig
   
   reasoner = Agent(
       name="Reasoner",
       llm="deepseek/deepseek-reasoner",
       output=OutputConfig(reasoning_steps=True),
   )
   extractor = Agent(name="Extractor", llm="gpt-4o-mini")
   
   AgentTeam(
       agents=[reasoner, extractor],
       tasks=[
           Task(description="Count r's in 'Strawberry'", agent=reasoner),
           Task(description="Give the final count only", agent=extractor),
       ],
   ).start()

How It Works

The reasoner surfaces chain-of-thought via OutputConfig(reasoning_steps=True), sequential tasks pass its output forward, and the extractor returns a short answer.

Use a reasoning model (e.g. deepseek-reasoner, o1-mini) for the first agent and a fast model for extraction.

Configuration Options

Best Practices

Related