Skip to content

Adaptive reasoning agent

GitHubでこのAIアクセラレーターにアクセス

This accelerator showcases an agent's ability to adapt its reasoning behavior based on conversation dynamics. The agent acts as a customer support representative for the "DataInsight Pro" analytics platform.

In this example, the agent uses:

  • GPT-4o for complex reasoning when corrections are detected.
  • GPT-4o-mini for fast responses during smooth conversation flow.
  • GPT-4o-mini as a reflection model to analyze the last three conversation turns and detect user corrections.

Specifically, the agent dynamically switches between models based on conversation analysis:

Scenario Model used 動作
Conversation flowing smoothly GPT-4o-mini Fast, direct responses
User corrects the agent GPT-4o More thorough reasoning
User rephrases question GPT-4o Agent recognizes confusion
Positive feedback received GPT-4o-mini Returns to efficient mode

The following diagram provides an overview that illustrates the agent architecture.

┌─────────────────────────────────────────────────────────────┐
│                     Frontend (React)                        │
│  ┌─────────────┐  ┌──────────────────────────────────────┐  │
│  │ Model Mode  │  │      Reflection Log Panel            │  │
│  │  Indicator  │  │  (shows gpt-4o-mini reasoning)       │  │
│  └─────────────┘  └──────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────┘
                              │
┌─────────────────────────────────────────────────────────────┐
│                   Adaptive Agent                            │
│  1. Store conversation history (last 3 turns)               │
│  2. Call reflection model (gpt-4o-mini) before response     │
│  3. Switch model based on correction detection              │
│     - Corrections detected → GPT-4o (thorough)              │
│     - Smooth conversation → GPT-4o-mini (fast)              │
└─────────────────────────────────────────────────────────────┘ 

The following script, with model adaptation noted, is applied in the accelerator:

Turn ユーザープロンプト Expected behavior モデル
1 "What pricing plans do you offer?" Lists 3 tiers (Starter, Pro, Enterprise) GPT-4o-mini
2 "How do I export data?" General export explanation GPT-4o-mini
3 "No, I meant export to CSV specifically, not PDF" Correction detected! Detailed CSV instructions GPT-4o
4 "Can I schedule automated exports?" Thorough answer with plan requirements GPT-4o
5 "Thanks, that's helpful!" Positive acknowledgment GPT-4o-mini