Skip to content
  • AI Development & Automation tags:
  • Gemini
  • AI
  • Agent
  • API keywords:
  • agent development
  • thinking level
  • API implementation guide

Gemini 3 Agent Development Guide: Implementing thinking_level and Autonomous Execution

In November 2025, Google launched an "agent-first" design. The key feature is controlling reasoning depth with the thinking_level parameter.

Target Audience

  • Intermediate developers interested in agent development with API

Key Points

  1. thinking_level parameter usage criteria
  2. thoughtSignature handling implementation
  3. OSS framework integration procedures

thinking_level Parameter Basics

The thinking_level parameter controls reasoning depth with two settings: low and high.

SettingUse CaseCharacteristics
lowSimple instruction followingMinimal latency, cost reduction
high (default)Deep planningMaximizes reasoning depth

thinking_level Implementation Example

import google.generativeai as genai

model = genai.GenerativeModel("gemini-3-pro")
response = model.generate_content(
    "Solve this complex math problem step by step",
    generation_config={"thinking_level": "high"}
)

By specifying thinking_level: "high", the model internally executes multi-layer reasoning for step-by-step problem solving.

thoughtSignature Handling

Capturing thoughtSignature is mandatory for Function Calling.

response = model.generate_content(
    "Get weather and suggest plan",
    tools=[weather_tool]
)

next_response = model.generate_content(
    "Revise previous suggestion",
    thought_signature=response.thought_signature
)

OSS Framework Integration

LangChain, LlamaIndex, and Pydantic AI are supported from day one.

LangChain Integration

from langchain_google_genai import ChatGoogleGenerativeAI
from langchain.agents import initialize_agent

llm = ChatGoogleGenerativeAI(model="gemini-3-pro")
agent = initialize_agent([search_tool], llm)
result = agent.run("Find Tokyo's population")

Issues and Solutions

SymptomSolution
Shallow reasoningExplicitly set "high"
Calling disconnectsPass thoughtSignature
Cost increaseUse low for simple tasks

Summary

Understanding thinking_level and thoughtSignature handling enables efficient agent development. Leverage frameworks like LangChain to build autonomous execution systems.

References