AI Innovations and Insights 30: Agentic Reasoning and Key Features of Claude 3.7 Sonnet

This article is the 30th in this deeply interesting series. In this post, we will explore two mind-opening topics:

• Agentic Reasoning: The Path to a Wise Researcher

• Claude 3.7 Sonnet: Thinking Budget Mechanism and Test-Time Compute Paradigm

Agentic Reasoning: The Path to a Wise Deep Researcher

Open-source code: https://github.com/theworldofagents/Agentic-Reasoning

Vivid Description

As shown in Figure 1, Agent Reasoning works like a skilled researcher. When tackling a problem, it generates the reasoning idea iteratively (LLM Reasoning), searches the web for information (Web Search Agent), writes code for computations (Code Agent), and organizes its thoughts using a mind map (Mind Map Agent) to clarify complex logical relationships.

Figure 1: The overall workflow of Agentic Reasoning. [Source].

Whenever it encounters a challenging question, it pauses to call on various "tool assistants" for answers or analysis. Finally, it brings all the information together, forming a complete reasoning chain to deliver an accurate and logically sound conclusion.

Overview

Deep Research has become a trending topic. It refers to intelligent agents capable of leveraging reasoning skills to integrate large volumes of online information and complete complex, multi-step research tasks.

Google, OpenAI, and Perplexity have each released their own versions of Deep Research, but none have made them open-source.

Agentic Reasoning is an open-source framework designed to enhance the reasoning abilities of LLMs by integrating external tools such as web search, code execution, and mind maps. It has outperformed Gemini Deep Research in deep research tasks, demonstrating strong performance across specialized fields such as finance, medicine, and law.

Here are the key agents in the Agentic Reasoning framework: