In this tutorial, we will explore how to create complex self-improvement AI agents using Google’s pointed-toe Gemini API. This self-improvement agent demonstrates autonomous problem solving, dynamically assessing performance, learning from success and failure, and iteratively enhances its abilities through reflective analysis and self-healing. This tutorial will introduce structured code implementations and details the mechanisms of memory management, function tracking, iterative task analysis, solution generation and performance evaluation, all integrated into a powerful self-learning feedback loop.
import google.generativeai as genai
import json
import time
import re
from typing import Dict, List, Any
from datetime import datetime
import tracebackWe built the basic components to leverage Google’s Generative AI API to build an AI-driven self-improvement proxy. Library such as JSON, TIME, RE, and DATETIME promotes structured data management, performance tracking, and text processing, while type prompts (dict, list, any) help ensure reliable and maintainable code.
class SelfImprovingAgent:
def __init__(self, api_key: str):
"""Initialize the self-improving agent with Gemini API"""
genai.configure(api_key=api_key)
self.model = genai.GenerativeModel('gemini-1.5-flash')
self.memory = {
'successful_strategies': [],
'failed_attempts': [],
'learned_patterns': [],
'performance_metrics': [],
'code_improvements': []
}
self.capabilities = {
'problem_solving': 0.5,
'code_generation': 0.5,
'learning_efficiency': 0.5,
'error_handling': 0.5
}
self.iteration_count = 0
self.improvement_history = []
def analyze_task(self, task: str) -> Dict[str, Any]:
"""Analyze a given task and determine approach"""
analysis_prompt = f"""
Analyze this task and provide a structured approach:
Task: {task}
Please provide:
1. Task complexity (1-10)
2. Required skills
3. Potential challenges
4. Recommended approach
5. Success criteria
Format as JSON.
"""
try:
response = self.model.generate_content(analysis_prompt)
json_match = re.search(r'{.*}', response.text, re.DOTALL)
if json_match:
return json.loads(json_match.group())
else:
return {
"complexity": 5,
"skills": ["general problem solving"],
"challenges": ["undefined requirements"],
"approach": "iterative improvement",
"success_criteria": ["task completion"]
}
except Exception as e:
print(f"Task analysis error: {e}")
return {"complexity": 5, "skills": [], "challenges": [], "approach": "basic", "success_criteria": []}
def solve_problem(self, problem: str) -> Dict[str, Any]:
"""Attempt to solve a problem using current capabilities"""
self.iteration_count += 1
print(f"n=== Iteration {self.iteration_count} ===")
print(f"Problem: {problem}")
task_analysis = self.analyze_task(problem)
print(f"Task Analysis: {task_analysis}")
solution_prompt = f"""
Based on my previous learning and capabilities, solve this problem:
Problem: {problem}
My current capabilities: {self.capabilities}
Previous successful strategies: {self.memory['successful_strategies'][-3:]} # Last 3
Known patterns: {self.memory['learned_patterns'][-3:]} # Last 3
Provide a detailed solution with:
1. Step-by-step approach
2. Code implementation (if applicable)
3. Expected outcome
4. Potential improvements
"""
try:
start_time = time.time()
response = self.model.generate_content(solution_prompt)
solve_time = time.time() - start_time
solution = {
'problem': problem,
'solution': response.text,
'solve_time': solve_time,
'iteration': self.iteration_count,
'task_analysis': task_analysis
}
quality_score = self.evaluate_solution(solution)
solution['quality_score'] = quality_score
self.memory['performance_metrics'].append({
'iteration': self.iteration_count,
'quality': quality_score,
'time': solve_time,
'complexity': task_analysis.get('complexity', 5)
})
if quality_score > 0.7:
self.memory['successful_strategies'].append(solution)
print(f"✅ Solution Quality: {quality_score:.2f} (Success)")
else:
self.memory['failed_attempts'].append(solution)
print(f"❌ Solution Quality: {quality_score:.2f} (Needs Improvement)")
return solution
except Exception as e:
print(f"Problem solving error: {e}")
error_solution = {
'problem': problem,
'solution': f"Error occurred: {str(e)}",
'solve_time': 0,
'iteration': self.iteration_count,
'quality_score': 0.0,
'error': str(e)
}
self.memory['failed_attempts'].append(error_solution)
return error_solution
def evaluate_solution(self, solution: Dict[str, Any]) -> float:
"""Evaluate the quality of a solution"""
evaluation_prompt = f"""
Evaluate this solution on a scale of 0.0 to 1.0:
Problem: {solution['problem']}
Solution: {solution['solution'][:500]}... # Truncated for evaluation
Rate based on:
1. Completeness (addresses all aspects)
2. Correctness (logically sound)
3. Clarity (well explained)
4. Practicality (implementable)
5. Innovation (creative approach)
Respond with just a decimal number between 0.0 and 1.0.
"""
try:
response = self.model.generate_content(evaluation_prompt)
score_match = re.search(r'(d+.?d*)', response.text)
if score_match:
score = float(score_match.group(1))
return min(max(score, 0.0), 1.0)
return 0.5
except:
return 0.5
def learn_from_experience(self):
"""Analyze past performance and improve capabilities"""
print("n🧠 Learning from experience...")
if len(self.memory['performance_metrics']) str:
"""Generate improved version of code"""
improvement_prompt = f"""
Improve this code based on the goal:
Current Code:
{current_code}
Improvement Goal: {improvement_goal}
My current capabilities: {self.capabilities}
Learned patterns: {self.memory['learned_patterns'][-3:]}
Provide improved code with:
1. Enhanced functionality
2. Better error handling
3. Improved efficiency
4. Clear comments explaining improvements
"""
try:
response = self.model.generate_content(improvement_prompt)
improved_code = {
'original': current_code,
'improved': response.text,
'goal': improvement_goal,
'iteration': self.iteration_count
}
self.memory['code_improvements'].append(improved_code)
return response.text
except Exception as e:
print(f"Code improvement error: {e}")
return current_code
def self_modify(self):
"""Attempt to improve the agent's own code"""
print("n🔧 Attempting self-modification...")
current_method = """
def solve_problem(self, problem: str) -> Dict[str, Any]:
# Current implementation
pass
"""
improved_method = self.generate_improved_code(
current_method,
"Make problem solving more efficient and accurate"
)
print("Generated improved method structure")
print("Note: Actual self-modification requires careful implementation in production")
def run_improvement_cycle(self, problems: List[str], cycles: int = 3):
"""Run a complete improvement cycle"""
print(f"🚀 Starting {cycles} improvement cycles with {len(problems)} problems")
for cycle in range(cycles):
print(f"n{'='*50}")
print(f"IMPROVEMENT CYCLE {cycle + 1}/{cycles}")
print(f"{'='*50}")
cycle_results = []
for problem in problems:
result = self.solve_problem(problem)
cycle_results.append(result)
time.sleep(1)
self.learn_from_experience()
if cycle str:
"""Generate a comprehensive performance report"""
if not self.memory['performance_metrics']:
return "No performance data available yet."
metrics = self.memory['performance_metrics']
avg_quality = sum(m['quality'] for m in metrics) / len(metrics)
avg_time = sum(m['time'] for m in metrics) / len(metrics)
report = f"""
📈 AGENT PERFORMANCE REPORT
{'='*40}
Total Iterations: {self.iteration_count}
Average Solution Quality: {avg_quality:.3f}
Average Solve Time: {avg_time:.2f}s
Successful Solutions: {len(self.memory['successful_strategies'])}
Failed Attempts: {len(self.memory['failed_attempts'])}
Success Rate: {len(self.memory['successful_strategies']) / max(1, self.iteration_count) * 100:.1f}%
Current Capabilities:
{json.dumps(self.capabilities, indent=2)}
Patterns Learned: {len(self.memory['learned_patterns'])}
Code Improvements: {len(self.memory['code_improvements'])}
"""
return reportWe define the above category as implementing a powerful framework that utilizes Google’s Gemini API for autonomous task resolution, self-evaluation, and adaptive learning. It combines structured memory systems, capability tracking, iterative problem solving, and continuous improvement cycles, even attempting to control self-modification. This advanced implementation allows agents to gradually improve their accuracy, efficiency and problem-solving sophistication, thereby creating a dynamic AI that can be independently developed and adapted.
def main():
"""Main function to demonstrate the self-improving agent"""
API_KEY = "Use Your GEMINI KEY Here"
if API_KEY == "Use Your GEMINI KEY Here":
print("⚠️ Please set your Gemini API key in the API_KEY variable")
print("Get your API key from:
return
agent = SelfImprovingAgent(API_KEY)
test_problems = [
"Write a function to calculate the factorial of a number",
"Create a simple text-based calculator that handles basic operations",
"Design a system to find the shortest path between two points in a graph",
"Implement a basic recommendation system for movies based on user preferences",
"Create a machine learning model to predict house prices based on features"
]
print("🤖 Self-Improving Agent Demo")
print("This agent will attempt to solve problems and improve over time")
agent.run_improvement_cycle(test_problems, cycles=3)
print("n" + agent.get_performance_report())
print("n" + "="*50)
print("TESTING IMPROVED AGENT")
print("="*50)
final_problem = "Create an efficient algorithm to sort a large dataset"
final_result = agent.solve_problem(final_problem)
print(f"nFinal Problem Solution Quality: {final_result.get('quality_score', 0):.2f}")
The main() function is the entry point to demonstrate the self-growth category class. It initializes the proxy using the user’s Gemini API keys and defines practical programming and system design tasks. The agent then iterates over these tasks, analyzing its performance to optimize its problem-solving capabilities to solve the problem through multiple improvement cycles. Finally, it tests agent enhancements with new complex tasks, demonstrates measurable progress and provides detailed performance reports.
def setup_instructions():
"""Print setup instructions for Google Colab"""
instructions = """
📋 SETUP INSTRUCTIONS FOR GOOGLE COLAB:
1. Install the Gemini API client:
!pip install google-generativeai
2. Get your Gemini API key:
- Go to
- Create a new API key
- Copy the key
3. Replace 'your-gemini-api-key-here' with your actual API key
4. Run the code!
🔧 CUSTOMIZATION OPTIONS:
- Modify test_problems list to add your own challenges
- Adjust improvement cycles count
- Add new capabilities to track
- Extend the learning mechanisms
💡 IMPROVEMENT IDEAS:
- Add persistent memory (save/load agent state)
- Implement more sophisticated evaluation metrics
- Add domain-specific problem types
- Create visualization of improvement over time
"""
print(instructions)
if __name__ == "__main__":
setup_instructions()
print("n" + "="*60)
main()
Finally, we define the setup_instructions() function, which guides users to prepare their Google Colab environment to run a self-improvement proxy. It explains how to install dependencies, set up and configure Gemini API keys, and highlights the various options to customize and enhance the proxy. This approach simplifies user onboarding, facilitates simple experiments and further expands the capabilities of the agent.
In summary, the implementation proven in this tutorial provides a comprehensive framework for creating AI agents that perform tasks and actively enhance their functionality over time. By leveraging the advanced generative power of the Gemini API and integrating structured self-improvement cycles, developers can build agents that have complex reasoning, iterative learning, and self-modification.
View notebooks on Github. All credits for this study are to the researchers on the project. Also, please stay tuned for us twitter And don’t forget to join us 95k+ ml reddit And subscribe Our newsletter.
Asif Razzaq is CEO of Marktechpost Media Inc. As a visionary entrepreneur and engineer, ASIF is committed to harnessing the potential of artificial intelligence to achieve social benefits. His recent effort is to launch Marktechpost, an artificial intelligence media platform that has an in-depth coverage of machine learning and deep learning news that can sound both technically, both through technical voices and be understood by a wide audience. The platform has over 2 million views per month, demonstrating its popularity among its audience.
