	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Agent Self-Improvement with Feedback Loops

ADK Agent Self-Improvement with Feedback Loops

Author: Venkata Sudhakar

Agent self-improvement uses feedback signals - thumbs up/down, explicit ratings, or correction messages - to refine the agent instruction over time. Rather than retraining the model, a meta-agent periodically analyses accumulated feedback and rewrites the system prompt to address recurring failure patterns. ShopMax India uses this to iteratively improve their product recommendation agent based on customer acceptance rates.

The pattern has three components: a feedback collector that stores ratings and corrections in Firestore, a feedback analyser that runs weekly and identifies patterns in negative feedback, and a prompt updater that rewrites the agent instruction based on the analysis.

The below example shows how ShopMax India implements the feedback collection and prompt refinement cycle.

from google.cloud import firestore
import google.generativeai as genai
from datetime import datetime

db = firestore.Client(project="shopmax-india")
genai.configure(api_key="your-api-key")
analyser = genai.GenerativeModel("gemini-2.0-flash")

def record_feedback(session_id: str, user_query: str,
                    agent_response: str, rating: int, correction: str = ""):
    """Record feedback: rating 1-5, optional correction text."""
    db.collection("agent_feedback").add({
        "session_id": session_id,
        "query": user_query,
        "response": agent_response[:500],
        "rating": rating,
        "correction": correction,
        "timestamp": datetime.utcnow().isoformat()
    })
    print(f"Feedback recorded: rating={rating}")

# Simulate feedback collection
record_feedback("sess_001",
    "Best phone under Rs 15,000",
    "I recommend Redmi Note 13 at Rs 14,999",
    rating=2,
    correction="Should mention EMI options and warranty details")

record_feedback("sess_002",
    "Good laptop for students",
    "Consider the ASUS VivoBook 15 at Rs 45,990",
    rating=2,
    correction="Student budgets are usually Rs 30,000-40,000, not Rs 45,000+")

The weekly prompt refinement job that analyses feedback and updates the agent,

def refine_agent_prompt(current_prompt: str) -> str:
    # Load recent negative feedback (rating < 3)
    feedback_docs = db.collection("agent_feedback")\
        .where("rating", "<", 3).limit(50).stream()
    negatives = [
        f"Query: {d.to_dict()['query']}\nIssue: {d.to_dict()['correction']}"
        for d in feedback_docs
    ]
    if not negatives:
        return current_prompt

analysis_prompt = f"""Current agent instruction:
{current_prompt}

Recent customer complaints:
{chr(10).join(negatives)}

Rewrite the instruction to fix these recurring issues.
Keep all existing rules. Add specific guidance to prevent the failures above.
Return only the improved instruction text."""

response = analyser.generate_content(analysis_prompt)
    new_prompt = response.text.strip()

# Save updated prompt
    db.collection("agent_config").document("recommendation_agent").set({
        "prompt": new_prompt,
        "updated_at": datetime.utcnow().isoformat(),
        "feedback_count": len(negatives)
    })
    print(f"Prompt refined using {len(negatives)} negative feedbacks")
    return new_prompt

# Run weekly
base_prompt = "You are a product recommendation agent for ShopMax India."
improved = refine_agent_prompt(base_prompt)
print("New prompt:", improved[:150])

It gives the following output,

Feedback recorded: rating=2
Feedback recorded: rating=2

Prompt refined using 2 negative feedbacks
New prompt: You are a product recommendation agent for ShopMax India.
Always mention EMI options and warranty when recommending products.
For student budgets assume Rs 25,000-40,000 unless stated otherwise.

Schedule the refinement job weekly using Cloud Scheduler. Store prompt versions in Firestore with timestamps so you can roll back if a new prompt performs worse. Track the average rating per week in BigQuery to measure whether self-improvement is working - ShopMax India improved average ratings from 3.2 to 4.1 over 8 weeks of automated refinement.

Send your comments, suggestions or queries regarding this site to [email protected].