AI Agents for Beginners - 9. Metacognition in AI Agents

AI Agents for Beginners - 9. Metacognition in AI Agents

Source: microsoft/ai-agents-for-beginners - 09-metacognition

What is Metacognition?

Importance of Metacognition
Mermaid
flowchart LR
    MC["Metacognition\nRecognizing Own Thought Process"]
    MC --> SR["Self-Reflection\nEvaluating Performance & Identifying Improvement Areas"]
    MC --> AD["Adaptability\nAdjusting Strategy Based on Past Experience"]
    MC --> EC["Error Correction\nAutonomous Error Detection & Correction"]
    MC --> RM["Resource Management\nOptimizing Time & Compute Resources"]

• "I prioritized cheap flights... but I might have missed direct flights. Let me check again."
• "Whenever the user said 'it's too complicated,' the issue was that I recommended items based on popularity."

Planning in Agents

Travel Agent Planning Flow
Mermaid
flowchart TD
    P1["1. Gather Preferences\nCollect dates, budget, interests"]
    --> P2["2. Retrieve Information\nSearch flights, hotels, attractions"]
    --> P3["3. Generate Recommendations\nCreate personalized itinerary"]
    --> P4["4. Present to User\nShare itinerary"]
    --> P5["5. Collect Feedback\nGather user feedback"]
    --> P6["6. Adjust Based on Feedback\nUpdate itinerary"]
    --> P7["7. Final Confirmation\nFinal review"]
    --> P8["8. Book & Confirm\nProceed with booking"]
    --> P9["9. Ongoing Support\nSupport during trip"]

travel_agent_metacognition.py
python

Metacognition Loop
Mermaid
flowchart TD
    G["gather_preferences()\nInitialize preferences"]
    --> R["generate_recommendations()\nSearch flights/hotels/attractions + Generate itinerary"]
    --> U["Present itinerary to user"]
    --> FB["User Feedback\n{liked, disliked}"]
    FB --> Adj["adjust_based_on_feedback()\nAccumulate experience_data\nReadjust preferences"]
    Adj -->|"Updated Preferences"| R

Corrective RAG System

RAG Tool Calling vs Pre-Emptive Context Load
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    subgraph RAG["RAG Tool Calling"]
        direction TD
        LLM1["LLM"]
        LLM1 --> T1["Tool"]
        LLM1 --> T2["Tool"]
        LLM1 --> T3["Tool"]
        T1 & T2 & T3 --> R1["Response"]
    end

    subgraph PCL["Pre-Emptive Context Load"]
        direction TD
        LLM2["LLM"]
        LLM2 --> CTX["Context\n(Provided before query processing)"]
        CTX --> R2["Response"]
    end

Corrective RAG Approach

Step-by-Step Implementation of Corrective RAG

step1_preferences.py
python

step2_retrieve.py
python

step3_feedback.py
python

• Prompting Technique: Formulate a new search query reflecting the feedback

step4_prompting.py
python

• Tool: Re-search with updated preferences and generate a new itinerary

step4_tool.py
python

• Evaluation: Continuous evaluation function reflecting feedback into preferences

step4_evaluation.py
python

corrective_rag_travel_agent.py
python

Corrective RAG Travel Agent Flow
Mermaid
flowchart TD
    GP["gather_preferences()\nStore destination, budget, interests"]
    --> GR["generate_recommendations()\nCall retrieve_information()"]
    GR --> RT["search_flights + hotels + attractions"]
    RT --> CI["create_itinerary()"]
    CI --> SU["Present itinerary to user"]
    SU --> FB["User Feedback\n{liked, disliked}"]
    FB --> AF["adjust_based_on_feedback()\nexperience_data.append(feedback)\nadjust_preferences() → Update preferences"]
    AF -->|"Updated user_preferences"| GR

Corrective RAG Flow

Corrective RAG Flow
Mermaid
flowchart TD
    User["User Input\nDestination, budget, interests"]
    --> Retrieve["Retrieve Information\nFlights, hotels, attractions"]
    --> Gen["Generate Initial Itinerary"]
    --> Feedback["Collect User Feedback"]
    Feedback --> Adjust{Analyze Feedback}
    Adjust -->|"disliked items"| NewQuery["Rewrite Query\nAdd preferences['avoid']"]
    NewQuery --> Retrieve
    Adjust -->|"Satisfied"| Done["Finalize Itinerary"]

Pre-emptive Context Load

preemptive_context.py
python

Pre-emptive Context Load Flow
Mermaid
flowchart TD
    Init["__init__()\nPre-load self.context\n{Paris, Tokyo, NY, Sydney}"]
    --> Query["User destination query"]
    --> Get["get_destination_info(destination)\nself.context.get(destination)"]
    Get --> Check{Context exists?}
    Check -->|"Yes"| Hit["country, currency, language, attractions\nReturn immediately (no external retrieval)"]
    Check -->|"No"| Miss["Return 'No information available'"]

Bootstrapping a Plan with a Goal Before Iterating

bootstrap_plan.py
python

Goal Bootstrapping Loop
Mermaid
flowchart TD
    Goal["Define Clear Goal\n(e.g., maximize satisfaction, budget ≤ $2000)"]
    --> Bootstrap["bootstrap_plan\nCreate initial plan based on goal & budget"]
    --> Execute["Execute Plan"]
    --> Eval{Evaluate Results}
    Eval -->|"Can be improved"| Iterate["iterate_plan\nReplace with better destinations"]
    Iterate --> Execute
    Eval -->|"Optimal"| Done["Finalize Travel Plan"]

Taking advantage of LLM for Re-ranking and Scoring

llm_reranking.py
python

LLM Re-ranking & Scoring Flow
Mermaid
flowchart TD
    Query["User Preferences\nactivity: sightseeing\nculture: diverse"]
    --> Retrieve["Initial Candidate List\n(Paris, Tokyo, NYC, Sydney)"]
    --> Prompt["generate_prompt()\nInclude preferences + destination descriptions"]
    --> LLM["LLM (Azure OpenAI)\nEvaluate relevance & quality"]
    LLM --> Rerank["Re-ranking\nSort best-matching destinations to the top"]
    Rerank --> Score["Scoring\nAssign scores to each candidate"]
    Score --> Result["Final Recommendations\n(Including ranks & scores)"]

RAG: Prompting Technique vs Tool

rag_prompting_vs_tool.py
python

Evaluating Relevancy

relevancy_techniques.py
python

travel_agent_relevancy.py
python

Relevancy Evaluation Flow
Mermaid
flowchart TD
    Pref["User Preferences\n{destination, budget, interests}"]
    --> GenRec["generate_recommendations()\nStart itinerary generation"]
    GenRec --> Retrieve["retrieve_information()\nSearch flights, hotels, attractions"]
    Retrieve --> Rank["filter_and_rank(hotels)\nSort by relevance_score → Top 10"]
    Rank --> Score["relevance_score(item, query)\nScore matching category, price, location"]
    Score --> Itin["create_itinerary()\nGenerate final itinerary"]
    Itin --> FB["User Feedback\n{liked, disliked}"]
    FB --> Adjust["adjust_based_on_feedback()\nAdjust relevance +1 / -1"]
    Adjust -->|"Reflect in next recommendation"| Rank

Search with Intent

search_with_intent.py
python

Intent-Aware Search Flow
Mermaid
zenuml
    title Intent-Aware Search Flow
    User->Agent: "best museums in Paris"
    Agent->IntentClassifier: identify_intent(query)
    IntentClassifier->Agent: "informational"
    Agent->ContextAnalyzer: analyze_context(query, user_history)
    ContextAnalyzer->Agent: current + history context
    Agent->SearchEngine: search informational query with preferences
    SearchEngine->Agent: raw results
    Agent->Personalizer: personalize_results(results, user_history)
    Personalizer->Agent: filtered top 10
    Agent->User: personalized recommendations

Generating Code as a Tool

step1_agent.py
python

step2_generate_code.py
python

step3_execute.py
python

step4_itinerary.py
python

step5_feedback_regen.py
python

Code Generating Agent Flow
Mermaid
flowchart TD
    S1["Step 1\ngather_preferences()\nCollect user preferences"]
    --> S2["Step 2\ngenerate_code_to_fetch_data()\ngenerate_code_to_fetch_hotels()\nDynamically generate code"]
    --> S3["Step 3\nexecute_code()\nExecute generated code → flights, hotels"]
    --> S4["Step 4\ngenerate_itinerary()\nCombine flights, hotels, attractions"]
    --> S5["Step 5\nadjust_based_on_feedback()\nUpdate preferences → Regenerate & re-execute code"]
    S5 -->|"Updated Preferences"| S2

Leveraging Environmental Awareness and Reasoning

schema_awareness.py
python

Schema-Aware Feedback Adjustment Flow
Mermaid
flowchart TD
    FB["User Feedback\n{liked, disliked}"]
    SC["Schema\n{favorites·avoid: +/-/default}"]
    FB & SC --> AdjFn["adjust_based_on_feedback()\nDirectly update favorites & avoid"]
    AdjFn --> Loop["Iterate over schema fields"]
    Loop --> EnvFn["adjust_based_on_environment()\nDetermine adjustment value per field"]
    EnvFn --> UPref["Updated Preferences"]
    UPref --> GenF["generate_code_to_fetch_data()"]
    UPref --> GenH["generate_code_to_fetch_hotels()"]
    GenF --> ExF["execute_code() → flights"]
    GenH --> ExH["execute_code() → hotels"]
    ExF & ExH --> Itin["generate_itinerary()\nUpdated Itinerary"]

Using SQL as a Retrieval-Augmented Generation (RAG) Technique

sql_as_rag.py
python

example_queries.sql
sql

SQL as RAG: Query Generation & Execution Flow
Mermaid
flowchart TD
    Pref["User Preferences\n{destination, budget, dates, interests}"]
    --> GenRec["generate_recommendations(preferences)"]
    GenRec --> FQ["generate_sql_query('flights')\nSELECT * FROM flights WHERE ..."]
    GenRec --> HQ["generate_sql_query('hotels')\nSELECT * FROM hotels WHERE ..."]
    GenRec --> AQ["generate_sql_query('attractions')\nSELECT * FROM attractions WHERE ..."]
    FQ --> EF["execute_sql_query()\n→ flight results"]
    HQ --> EH["execute_sql_query()\n→ hotel results"]
    AQ --> EA["execute_sql_query()\n→ attraction results"]
    EF & EH & EA --> Dict["{flights, hotels, attractions}\nItinerary Dict"]

Example of Metacognition

hotel_metacognition.py
python

Metacognition: Reflect & Adjust Loop
Mermaid
flowchart TD
    Start["Initial Decision\nrecommend_hotel('cheapest')"]
    --> Eval["Self-Reflection\nreflect_on_choice()"]
    Eval --> Feedback{User Feedback}
    Feedback -->|"'bad'\nprice < 100 or quality < 7"| Switch["Switch Strategy\n'cheapest' → 'highest_quality'"]
    Switch --> Retry["Re-recommend\nrecommend_hotel('highest_quality')"]
    Retry --> Eval
    Feedback -->|"'good'"| Done["Confirm Recommendation"]

• Initial Decision: Selected Budget Inn with the 'cheapest' strategy
• Self-Reflection: Recognized "bad" feedback due to quality 6
• Strategy Adjustment: Ultimately recommended Luxury Stay

Summary

Lesson 09 Summary
Mermaid
flowchart LR
    Root["Metacognition\nin AI Agents"]
    Root --> MC["Metacognition Loop"]
    Root --> RAG["RAG Strategies"]
    Root --> Search["Intent Search"]
    Root --> Code["Code Generation"]
    MC --> M1["Self-Reflection\nPerformance Evaluation"]
    MC --> M2["Error Detection\nError Recognition"]
    MC --> M3["Strategy Adjustment\nModifying Strategy"]
    RAG --> R1["Corrective RAG\nFeedback-based Re-search"]
    RAG --> R2["Pre-emptive Load\nPre-load Context"]
    RAG --> R3["SQL as RAG\nDynamic Query Generation"]
    Search --> S1["Informational"]
    Search --> S2["Navigational"]
    Search --> S3["Transactional"]
    Code --> C1["generate_code\nDynamic Code Generation"]
    Code --> C2["execute_code\nExecution & Results Reflection"]

• Metacognition is the ability of an agent to recognize and modify its own decision-making process, representing an improvement of the reasoning strategy itself rather than a simple correction of outputs.
• Corrective RAG implements an iterative self-correction loop that incorporates user feedback into preferences/avoids, rewrites queries, and performs re-retrieval.
• Pre-emptive Context Load pre-loads known domain knowledge into memory to provide instant responses without external retrieval, making it highly effective when latency is critical.
• Intent-aware Search classifies queries into Informational, Navigational, or Transactional intents, delivering far more relevant results than simple keyword matching.
• Code Generation + SQL as RAG empowers agents to dynamically write and execute code to retrieve real-time data and automate complex tasks.