"""
Pattern 053: Mixture of Agents (MoA)
Description:
The Mixture of Agents pattern combines outputs from multiple specialized agents
using learned or rule-based aggregation strategies. Each agent brings specific
expertise, and their outputs are intelligently combined to produce superior
results that leverage the strengths of each individual agent.
Components:
1. Specialized Agents: Multiple agents with different capabilities/models
2. Proposer Agents: Generate initial candidate responses
3. Aggregator Agent: Combines and synthesizes multiple responses
4. Confidence Scoring: Weights agent contributions
5. Quality Evaluation: Assesses individual and combined outputs
Use Cases:
- Complex tasks requiring diverse expertise
- Multi-domain question answering
- Code generation with multiple approaches
- Creative content requiring varied perspectives
- Research synthesis from multiple sources
- Decision-making with multiple viewpoints
LangChain Implementation:
Uses multiple LLM instances with different configurations or models,
then aggregates their outputs using a meta-agent that synthesizes
the best aspects of each response.
"""
import os
import time
from typing import Dict, Any, List, Optional, Tuple
from dataclasses import dataclass, field
from datetime import datetime
from enum import Enum
from dotenv import load_dotenv
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_openai import ChatOpenAI
load_dotenv()
class AgentRole(Enum):
"""Roles for agents in MoA"""
PROPOSER = "proposer" # Generates candidate responses
AGGREGATOR = "aggregator" # Combines multiple responses
CRITIC = "critic" # Evaluates quality
SPECIALIST = "specialist" # Domain expert
class AggregationStrategy(Enum):
"""Strategies for combining agent outputs"""
SYNTHESIS = "synthesis" # Synthesize best parts
VOTING = "voting" # Majority vote or ranking
WEIGHTED = "weighted" # Weighted by confidence
SEQUENTIAL = "sequential" # Sequential refinement
BEST_OF = "best_of" # Select single best
@dataclass
class AgentOutput:
"""Output from a single agent"""
agent_id: str
agent_role: AgentRole
response: str
confidence: float # 0.0-1.0
reasoning: Optional[str] = None
metadata: Dict[str, Any] = field(default_factory=dict)
def to_dict(self) -> Dict[str, Any]:
return {
"agent_id": self.agent_id,
"role": self.agent_role.value,
"response": self.response[:150] + "..." if len(self.response) > 150 else self.response,
"confidence": f"{self.confidence:.2f}",
"reasoning": self.reasoning[:100] + "..." if self.reasoning and len(self.reasoning) > 100 else self.reasoning
}
@dataclass
class MoAResult:
"""Result from Mixture of Agents execution"""
query: str
final_response: str
aggregation_strategy: AggregationStrategy
agent_outputs: List[AgentOutput]
aggregator_reasoning: str
confidence_score: float
execution_time_ms: float
layer_count: int
@property
def num_agents(self) -> int:
return len(self.agent_outputs)
def to_dict(self) -> Dict[str, Any]:
return {
"query": self.query[:100] + "..." if len(self.query) > 100 else self.query,
"final_response": self.final_response[:200] + "..." if len(self.final_response) > 200 else self.final_response,
"strategy": self.aggregation_strategy.value,
"num_agents": self.num_agents,
"layer_count": self.layer_count,
"confidence": f"{self.confidence_score:.2f}",
"execution_time_ms": f"{self.execution_time_ms:.1f}"
}
class MixtureOfAgents:
"""
Mixture of Agents implementation with multiple aggregation strategies.
This implementation provides:
1. Multiple proposer agents generating diverse responses
2. Aggregator agent synthesizing best combined response
3. Layered architecture for iterative refinement
4. Confidence-based weighting
5. Multiple aggregation strategies
"""
def __init__(
self,
num_proposers: int = 3,
aggregation_strategy: AggregationStrategy = AggregationStrategy.SYNTHESIS,
temperature: float = 0.7,
use_layers: bool = True,
num_layers: int = 2
):
self.num_proposers = num_proposers
self.aggregation_strategy = aggregation_strategy
self.temperature = temperature
self.use_layers = use_layers
self.num_layers = num_layers if use_layers else 1
# Create proposer agents with varied temperatures for diversity
self.proposers = [
ChatOpenAI(
model_name="gpt-3.5-turbo",
temperature=temperature + (i * 0.1) # Vary temperature
)
for i in range(num_proposers)
]
# Aggregator agent (lower temperature for focused synthesis)
self.aggregator = ChatOpenAI(
model_name="gpt-3.5-turbo",
temperature=0.3
)
def _generate_proposer_response(
self,
agent_id: str,
llm: ChatOpenAI,
query: str,
context: Optional[str] = None
) -> AgentOutput:
"""Generate response from a proposer agent"""
prompt = ChatPromptTemplate.from_messages([
("system", """You are an expert assistant providing thoughtful, comprehensive answers.
Provide your best response to the user's question."""),
("user", "{query}")
])
if context:
prompt = ChatPromptTemplate.from_messages([
("system", """You are an expert assistant. You have access to previous responses from other agents.
Consider their insights but provide your own unique perspective and improvements.
Previous responses:
{context}"""),
("user", "Now provide your response to: {query}")
])
try:
chain = prompt | llm | StrOutputParser()
if context:
response = chain.invoke({"query": query, "context": context})
else:
response = chain.invoke({"query": query})
# Estimate confidence based on response length and coherence
confidence = min(0.9, 0.5 + (len(response) / 1000))
return AgentOutput(
agent_id=agent_id,
agent_role=AgentRole.PROPOSER,
response=response,
confidence=confidence
)
except Exception as e:
return AgentOutput(
agent_id=agent_id,
agent_role=AgentRole.PROPOSER,
response=f"Error: {str(e)}",
confidence=0.0,
metadata={"error": True}
)
def _aggregate_synthesis(
self,
query: str,
agent_outputs: List[AgentOutput]
) -> Tuple[str, str, float]:
"""Synthesize multiple responses into one cohesive answer"""
# Prepare responses for aggregator
responses_text = "\n\n".join([
f"Agent {i+1} (Confidence: {output.confidence:.2f}):\n{output.response}"
for i, output in enumerate(agent_outputs)
if not output.metadata.get("error")
])
synthesis_prompt = ChatPromptTemplate.from_messages([
("system", """You are an expert aggregator. You receive multiple responses to the same question.
Your task:
1. Identify the best insights from each response
2. Synthesize them into a single, comprehensive answer
3. Resolve any contradictions
4. Provide a cohesive, well-structured response
Be concise but thorough. Leverage the strengths of each response."""),
("user", """Question: {query}
Multiple Agent Responses:
{responses}
Synthesize these into the best possible answer:""")
])
chain = synthesis_prompt | self.aggregator | StrOutputParser()
try:
final_response = chain.invoke({
"query": query,
"responses": responses_text
})
# Calculate confidence as weighted average
valid_outputs = [o for o in agent_outputs if not o.metadata.get("error")]
avg_confidence = sum(o.confidence for o in valid_outputs) / len(valid_outputs) if valid_outputs else 0.0
reasoning = "Synthesized best insights from all agents"
return final_response, reasoning, avg_confidence
except Exception as e:
# Fallback to best individual response
best = max(agent_outputs, key=lambda x: x.confidence)
return best.response, f"Synthesis failed: {str(e)}", best.confidence
def _aggregate_voting(
self,
query: str,
agent_outputs: List[AgentOutput]
) -> Tuple[str, str, float]:
"""Use voting to select best response"""
# For voting, we'll use the aggregator to rank responses
responses_text = "\n\n".join([
f"Response {i+1}:\n{output.response}"
for i, output in enumerate(agent_outputs)
if not output.metadata.get("error")
])
voting_prompt = ChatPromptTemplate.from_messages([
("system", """You are evaluating multiple responses to select the best one.
Rank the responses and select the winner. Explain your choice briefly.
Format:
Winner: [Response number]
Reason: [Brief explanation]"""),
("user", """Question: {query}
Responses to evaluate:
{responses}
Select the best response:""")
])
chain = voting_prompt | self.aggregator | StrOutputParser()
try:
evaluation = chain.invoke({
"query": query,
"responses": responses_text
})
# Parse winner (simplified)
winner_idx = 0
for i in range(len(agent_outputs)):
if f"Response {i+1}" in evaluation or f"#{i+1}" in evaluation:
winner_idx = i
break
winner = agent_outputs[winner_idx]
return winner.response, evaluation, winner.confidence
except Exception as e:
best = max(agent_outputs, key=lambda x: x.confidence)
return best.response, f"Voting failed: {str(e)}", best.confidence
def _aggregate_weighted(
self,
query: str,
agent_outputs: List[AgentOutput]
) -> Tuple[str, str, float]:
"""Weighted combination based on confidence scores"""
# Similar to synthesis but explicitly weights by confidence
valid_outputs = [o for o in agent_outputs if not o.metadata.get("error")]
if not valid_outputs:
return "No valid responses", "All agents failed", 0.0
# Sort by confidence
sorted_outputs = sorted(valid_outputs, key=lambda x: x.confidence, reverse=True)
# Give more weight to high-confidence responses
weighted_text = "\n\n".join([
f"Agent {i+1} (High Confidence - Weight: {output.confidence:.2f}):\n{output.response}"
if output.confidence > 0.7 else
f"Agent {i+1} (Moderate Confidence - Weight: {output.confidence:.2f}):\n{output.response}"
for i, output in enumerate(sorted_outputs)
])
weighted_prompt = ChatPromptTemplate.from_messages([
("system", """You are combining multiple responses weighted by their confidence scores.
Give more importance to high-confidence responses but don't ignore lower-confidence ones
if they contain unique valuable insights.
Produce a balanced synthesis."""),
("user", """Question: {query}
Weighted Responses:
{responses}
Produce weighted synthesis:""")
])
chain = weighted_prompt | self.aggregator | StrOutputParser()
final_response = chain.invoke({
"query": query,
"responses": weighted_text
})
# Confidence is weighted average favoring high-confidence responses
total_weight = sum(o.confidence ** 2 for o in sorted_outputs)
weighted_confidence = sum(
(o.confidence ** 2) * o.confidence for o in sorted_outputs
) / total_weight if total_weight > 0 else 0.5
reasoning = "Weighted synthesis favoring high-confidence responses"
return final_response, reasoning, weighted_confidence
def _aggregate_sequential(
self,
query: str,
agent_outputs: List[AgentOutput]
) -> Tuple[str, str, float]:
"""Sequential refinement - each response builds on previous"""
# Start with first response
current_response = agent_outputs[0].response
for i, output in enumerate(agent_outputs[1:], 1):
if output.metadata.get("error"):
continue
refine_prompt = ChatPromptTemplate.from_messages([
("system", """You are refining an answer by incorporating new insights.
Current answer:
{current}
New perspective:
{new}
Produce an improved, refined answer that combines both:"""),
("user", "Refined answer:")
])
chain = refine_prompt | self.aggregator | StrOutputParser()
current_response = chain.invoke({
"current": current_response,
"new": output.response
})
avg_confidence = sum(o.confidence for o in agent_outputs) / len(agent_outputs)
reasoning = f"Sequential refinement through {len(agent_outputs)} agents"
return current_response, reasoning, avg_confidence
def execute(self, query: str) -> MoAResult:
"""Execute Mixture of Agents with multiple layers"""
start_time = time.time()
all_agent_outputs = []
current_context = None
# Execute layers
for layer in range(self.num_layers):
layer_outputs = []
# Generate responses from all proposers
for i, proposer in enumerate(self.proposers):
agent_id = f"layer{layer+1}_agent{i+1}"
output = self._generate_proposer_response(
agent_id,
proposer,
query,
context=current_context
)
layer_outputs.append(output)
all_agent_outputs.append(output)
# Prepare context for next layer
if layer < self.num_layers - 1:
valid_responses = [
o.response for o in layer_outputs
if not o.metadata.get("error")
]
current_context = "\n\n".join(valid_responses[:2]) # Top 2 responses
# Aggregate final responses
final_layer_outputs = all_agent_outputs[-self.num_proposers:]
if self.aggregation_strategy == AggregationStrategy.SYNTHESIS:
final_response, reasoning, confidence = self._aggregate_synthesis(
query, final_layer_outputs
)
elif self.aggregation_strategy == AggregationStrategy.VOTING:
final_response, reasoning, confidence = self._aggregate_voting(
query, final_layer_outputs
)
elif self.aggregation_strategy == AggregationStrategy.WEIGHTED:
final_response, reasoning, confidence = self._aggregate_weighted(
query, final_layer_outputs
)
elif self.aggregation_strategy == AggregationStrategy.SEQUENTIAL:
final_response, reasoning, confidence = self._aggregate_sequential(
query, final_layer_outputs
)
else: # BEST_OF
best = max(final_layer_outputs, key=lambda x: x.confidence)
final_response = best.response
reasoning = f"Selected best individual response (confidence: {best.confidence:.2f})"
confidence = best.confidence
execution_time_ms = (time.time() - start_time) * 1000
return MoAResult(
query=query,
final_response=final_response,
aggregation_strategy=self.aggregation_strategy,
agent_outputs=all_agent_outputs,
aggregator_reasoning=reasoning,
confidence_score=confidence,
execution_time_ms=execution_time_ms,
layer_count=self.num_layers
)
def demonstrate_mixture_of_agents():
"""Demonstrate Mixture of Agents pattern"""
print("=" * 80)
print("PATTERN 053: MIXTURE OF AGENTS (MoA) DEMONSTRATION")
print("=" * 80)
print("\nCombining multiple specialized agents for superior results\n")
# Test 1: Basic MoA with synthesis
print("\n" + "=" * 80)
print("TEST 1: Basic MoA with Synthesis Aggregation")
print("=" * 80)
moa1 = MixtureOfAgents(
num_proposers=3,
aggregation_strategy=AggregationStrategy.SYNTHESIS,
use_layers=False
)
query1 = "What are the pros and cons of remote work?"
print(f"\nQuery: {query1}")
print(f"Number of Proposer Agents: {moa1.num_proposers}")
print(f"Aggregation Strategy: {moa1.aggregation_strategy.value}")
result1 = moa1.execute(query1)
print(f"\nš Individual Agent Responses:")
for i, output in enumerate(result1.agent_outputs, 1):
print(f"\n Agent {i} (Confidence: {output.confidence:.2f}):")
print(f" {output.response[:120]}...")
print(f"\nāØ Final Synthesized Response:")
print(f" {result1.final_response[:300]}...")
print(f"\n Aggregator Reasoning: {result1.aggregator_reasoning}")
print(f" Final Confidence: {result1.confidence_score:.2f}")
print(f" Execution Time: {result1.execution_time_ms:.1f}ms")
# Test 2: Layered MoA
print("\n" + "=" * 80)
print("TEST 2: Layered MoA (Iterative Refinement)")
print("=" * 80)
moa2 = MixtureOfAgents(
num_proposers=3,
aggregation_strategy=AggregationStrategy.SYNTHESIS,
use_layers=True,
num_layers=2
)
query2 = "Explain quantum computing in simple terms"
print(f"\nQuery: {query2}")
print(f"Architecture: {moa2.num_layers} layers with {moa2.num_proposers} agents per layer")
result2 = moa2.execute(query2)
print(f"\nš Layer-by-Layer Generation:")
layer1_agents = result2.agent_outputs[:moa2.num_proposers]
layer2_agents = result2.agent_outputs[moa2.num_proposers:]
print(f"\n Layer 1 (Initial Responses):")
for i, output in enumerate(layer1_agents, 1):
print(f" Agent {i}: {output.response[:80]}...")
if layer2_agents:
print(f"\n Layer 2 (Refined with Context):")
for i, output in enumerate(layer2_agents, 1):
print(f" Agent {i}: {output.response[:80]}...")
print(f"\nāØ Final Layered Response:")
print(f" {result2.final_response[:300]}...")
print(f" Confidence: {result2.confidence_score:.2f}")
# Test 3: Different aggregation strategies
print("\n" + "=" * 80)
print("TEST 3: Comparing Aggregation Strategies")
print("=" * 80)
query3 = "What's the best way to learn programming?"
strategies = [
AggregationStrategy.SYNTHESIS,
AggregationStrategy.VOTING,
AggregationStrategy.WEIGHTED
]
print(f"\nQuery: {query3}")
print(f"\nTesting different aggregation strategies...")
results = []
for strategy in strategies:
moa = MixtureOfAgents(
num_proposers=3,
aggregation_strategy=strategy,
use_layers=False
)
result = moa.execute(query3)
results.append((strategy, result))
print(f"\n {strategy.value.upper()}:")
print(f" Response: {result.final_response[:100]}...")
print(f" Confidence: {result.confidence_score:.2f}")
print(f" Reasoning: {result.aggregator_reasoning[:80]}...")
# Summary
print("\n" + "=" * 80)
print("MIXTURE OF AGENTS PATTERN SUMMARY")
print("=" * 80)
print("""
Key Benefits:
1. Diverse Perspectives: Multiple agents provide varied viewpoints
2. Error Reduction: Aggregation filters out individual mistakes
3. Robustness: Less sensitive to single agent failures
4. Quality Improvement: Combined output better than any individual
5. Expertise Combination: Leverage multiple specializations
Implementation Features:
1. Multiple proposer agents with diverse configurations
2. Aggregator agent for intelligent synthesis
3. Layered architecture for iterative refinement
4. Confidence-based weighting
5. Multiple aggregation strategies
Aggregation Strategies:
1. Synthesis: Combines best insights from all responses
2. Voting: Democratic selection of best response
3. Weighted: Confidence-based combination
4. Sequential: Iterative refinement chain
5. Best-of: Simple selection of highest quality
Layered Architecture:
- Layer 1: Initial independent responses
- Layer 2+: Responses informed by previous layer
- Each layer refines and improves outputs
- Context flows from layer to layer
Use Cases:
- Complex questions requiring multiple perspectives
- Creative tasks needing diverse ideas
- Technical problems with multiple solutions
- Research synthesis from various sources
- Decision-making with high stakes
- Content generation requiring variety
Best Practices:
1. Use 3-5 proposer agents (balance diversity vs cost)
2. Vary agent configurations for diversity
3. Match aggregation strategy to use case
4. Use layers for complex reasoning tasks
5. Monitor confidence scores for quality
6. A/B test different strategies
7. Cache common query results
Production Considerations:
- Parallel execution for latency (async)
- Cost vs quality tradeoffs
- Model diversity (different providers)
- Caching for repeated queries
- Load balancing across agents
- Timeout handling
- Fallback strategies
Comparison with Related Patterns:
- vs. Ensemble: Synthesis vs simple voting
- vs. Redundancy: Quality improvement vs fault tolerance
- vs. Multi-Agent Debate: Aggregation vs deliberation
- vs. Leader-Follower: Parallel vs hierarchical
The Mixture of Agents pattern is powerful for tasks requiring high quality
outputs that benefit from multiple perspectives and expert combination.
""")
if __name__ == "__main__":
demonstrate_mixture_of_agents()
Agentic Design Pattern: Mixture Of Agents
Pattern 053: Mixture Of Agents – Line-by-Line Explanation
Overview
This document provides a detailed line-by-line explanation of the Mixture Of Agents pattern implementation.
Pattern Description
Pattern 053: Mixture of Agents (MoA)
Description: The Mixture of Agents pattern combines outputs from multiple specialized agents using learned or rule-based aggregation strategies. Each agent brings specific expertise, and their outputs are intelligently combined to produce superior results that leverage the strengths of each individual agent.
Components: 1. Specialized Agents: Multiple agents with different capabilities/models 2. Proposer Agents: Generate initial candidate responses 3. Aggregator Agent: Combines and synthesizes multiple responses 4. Confidence Scoring: Weights agent contributions 5. Quality Evaluation: Assesses individual and combined outputs
Use Cases: – Complex tasks requiring diverse expertise – Multi-domain question answering – Code generation with multiple approaches – Creative content requiring varied perspectives – Research synthesis from multiple sources – Decision-making with multiple viewpoints
LangChain Implementation: Uses multiple LLM instances with different configurations or models, then aggregates their outputs using a meta-agent that synthesizes the best aspects of each response.
File Statistics
- Total Lines: 656
- Classes: 5
- Functions: 1
- Constants: 0
Module Documentation
"""
Pattern 053: Mixture of Agents (MoA)
Description:
The Mixture of Agents pattern combines outputs from multiple specialized agents
using learned or rule-based aggregation strategies. Each agent brings specific
expertise, and their outputs are intelligently combined to produce superior
results that leverage the strengths of each individual agent.
Components:
1. Specialized Agents: Multiple agents with different capabilities/models
2. Proposer Agents: Generate initial candidate responses
3. Aggregator Agent: Combines and synthesizes multiple responses
4. Confidence Scoring: Weights agent contributions
5. Quality Evaluation: Assesses individual and combined outputs
Use Cases:
- Complex tasks requiring diverse expertise
- Multi-domain question answering
- Code generation with multiple approaches
- Creative content requiring varied perspectives
- Research synthesis from multiple sources
- Decision-making with multiple viewpoints
LangChain Implementation:
Uses multiple LLM instances with different configurations or models,
then aggregates their outputs using a meta-agent that synthesizes
the best aspects of each response.
"""
Purpose: This module implements the Mixture Of Agents pattern, providing a structured approach to agent design and implementation.
Import Statements
Standard Library Imports
import dataclasses.dataclass
import dataclasses.field
import datetime.datetime
import enum.Enum
import langchain_core.output_parsers.StrOutputParser
import langchain_core.prompts.ChatPromptTemplate
import os
import time
import typing.Any
import typing.Dict
import typing.List
import typing.Optional
import typing.Tuple
Purpose: Core Python functionality:
– dataclasses.dataclass: Simplified class definitions with @dataclass
– dataclasses.field: Simplified class definitions with @dataclass
– datetime.datetime: Date and time handling
– enum.Enum: Enumeration support
– langchain_core.output_parsers.StrOutputParser: Regular expression operations
– langchain_core.prompts.ChatPromptTemplate: Regular expression operations
– os: Environment variable access and system operations
– time: Time-related functions
– typing.Any: Type hints for better code documentation (List, Dict, Any, Optional)
– typing.Dict: Type hints for better code documentation (List, Dict, Any, Optional)
– typing.List: Type hints for better code documentation (List, Dict, Any, Optional)
– typing.Optional: Type hints for better code documentation (List, Dict, Any, Optional)
– typing.Tuple: Type hints for better code documentation (List, Dict, Any, Optional)
LangChain Imports
from langchain_openai.ChatOpenAI
Purpose: LangChain framework components: – LLM Interface: Wrapper for OpenAI models – Chains: Composable building blocks for complex workflows
Other Imports
from dotenv.load_dotenv
Purpose: Additional dependencies for specialized functionality.
Classes
Class: AgentRole
Lines 46-51
Inherits from: Enum
Purpose:
Roles for agents in MoA
Class: AggregationStrategy
Lines 54-60
Inherits from: Enum
Purpose:
Strategies for combining agent outputs
Class: AgentOutput
Lines 64-80
Purpose:
Output from a single agent
Methods:
to_dict(self)
Lines 73-80
Purpose: To Dict operation.
Parameters:
Returns: <ast.Subscript object at 0x000001A06D68EF90>
Class: MoAResult
Lines 84-108
Purpose:
Result from Mixture of Agents execution
Methods:
num_agents(self)
Lines 96-97
Decorators: property
Purpose: Num Agents operation.
Parameters:
Returns: int
to_dict(self)
Lines 99-108
Purpose: To Dict operation.
Parameters:
Returns: <ast.Subscript object at 0x000001A06D17C750>
Class: MixtureOfAgents
Lines 111-475
Purpose:
Mixture of Agents implementation with multiple aggregation strategies.
This implementation provides:
1. Multiple proposer agents generating diverse responses
2. Aggregator agent synthesizing best combined response
3. Layered architecture for iterative refinement
4. Confidence-based weighting
5. Multiple aggregation strategies
Methods:
__init__(self, num_proposers, aggregation_strategy, temperature, use_layers, num_layers)
Lines 123-150
Purpose: Initialize the class instance with required parameters.
Parameters:
– num_proposers: Num proposers
– aggregation_strategy: Aggregation strategy
– temperature: Sampling temperature (0-1)
– use_layers: Use layers
– num_layers: Num layers
_generate_proposer_response(self, agent_id, llm, query, context)
Lines 152-202
Purpose:
Generate response from a proposer agent
Parameters:
– agent_id: Agent id
– llm: Language model
– query: Input query or question
– context: Context
Returns: AgentOutput
_aggregate_synthesis(self, query, agent_outputs)
Lines 204-255
Purpose:
Synthesize multiple responses into one cohesive answer
Parameters:
– query: Input query or question
– agent_outputs: Agent outputs
Returns: <ast.Subscript object at 0x000001A06D614950>
_aggregate_voting(self, query, agent_outputs)
Lines 257-308
Purpose:
Use voting to select best response
Parameters:
– query: Input query or question
– agent_outputs: Agent outputs
Returns: <ast.Subscript object at 0x000001A06D674A50>
_aggregate_weighted(self, query, agent_outputs)
Lines 310-364
Purpose:
Weighted combination based on confidence scores
Parameters:
– query: Input query or question
– agent_outputs: Agent outputs
Returns: <ast.Subscript object at 0x000001A06D73B0D0>
_aggregate_sequential(self, query, agent_outputs)
Lines 366-403
Purpose:
Sequential refinement - each response builds on previous
Parameters:
– query: Input query or question
– agent_outputs: Agent outputs
Returns: <ast.Subscript object at 0x000001A06D622590>
execute(self, query)
Lines 405-475
Purpose:
Execute Mixture of Agents with multiple layers
Parameters:
– query: Input query or question
Returns: MoAResult
Functions
demonstrate_mixture_of_agents()
Lines 478-651
Purpose:
Demonstrate Mixture of Agents pattern
Key Concepts Summary
Pattern Overview
The Mixture Of Agents pattern provides a structured approach to building AI agents with specific capabilities and behaviors. This implementation leverages LangChain’s composable components to create a robust and maintainable solution.
Core Components
- Agent Class: Main implementation of the pattern logic
- Prompt Engineering: Carefully crafted prompts to guide LLM behavior
- Tool Integration: Optional tools for extended functionality
- Memory Management: State handling across interactions
- Output Parsing: Structured response processing
When to Use
ā Good for: – Scenarios requiring this specific pattern’s strengths – Applications needing structured agent behavior – Systems where explainability is important – Production environments requiring reliability
ā Not ideal for: – Simple single-prompt tasks – Scenarios requiring different pattern characteristics – Resource-constrained environments (if computationally intensive)
Best Practices Demonstrated
1. Environment Configuration
- ā
Use
.envfiles for API keys - ā Never hardcode credentials
- ā Validate environment variables on startup
2. Error Handling
- ā Graceful degradation on failures
- ā Informative error messages
- ā Logging for debugging
3. Prompt Engineering
- ā Clear, specific instructions
- ā Structured output formats
- ā Few-shot examples when beneficial
4. Code Organization
- ā Separation of concerns (class-based design)
- ā Type hints for clarity
- ā Docstrings for documentation
5. Performance Optimization
- ā Appropriate temperature settings
- ā Token budget management
- ā Caching where applicable
LangChain Components Used
- ChatOpenAI: OpenAI model interface
- Chains: Sequential component composition
Why LangChain? – Composable abstractions for LLM applications – Rich ecosystem of integrations – Production-ready patterns – Active community and development
Execution Flow
1. Initialize agent with configuration
ā
2. Load environment and dependencies
ā
3. Prepare prompt with input data
ā
4. Invoke LLM through LangChain
ā
5. Parse and validate response
ā
6. Return formatted result
Key Decision Points: – Model selection (temperature, model type) – Prompt structure (few-shot vs zero-shot) – Error handling strategy – Output format validation
Performance Considerations
Computational Cost
- LLM Calls: Primary cost factor (based on tokens)
- Latency: Depends on model and input size
- Memory: Minimal for most patterns
Optimization Strategies
- Prompt Optimization: Minimize tokens while maintaining clarity
- Caching: Cache repeated queries/responses
- Model Selection: Use smaller models when appropriate
- Batch Processing: Group similar requests
- Streaming: For long responses, stream tokens
Cost Estimates (Approximate)
- GPT-3.5-turbo: ~$0.002 per request
- GPT-4: ~$0.03 per request
- GPT-4-turbo: ~$0.01 per request
Actual costs vary based on input/output token counts
Common Pitfalls to Avoid
1. Prompt Design Issues
ā Vague or ambiguous instructions
ā
Clear, specific, testable prompts
2. Error Handling
ā Ignoring API errors or rate limits
ā
Implement retry logic and graceful degradation
3. Token Management
ā Exceeding model token limits
ā
Truncate inputs, chunk processing
4. Security
ā Exposing API keys in code
ā
Use environment variables
5. Testing
ā Only testing happy paths
ā
Test edge cases, failures, malformed inputs
6. Cost Management
ā Unmonitored API usage
ā
Implement budgets, alerts, and monitoring
Further Improvements
Production Enhancements
- Monitoring: Add logging, metrics, tracing
- Validation: Input/output schema validation
- Rate Limiting: Implement request throttling
- Caching: Redis or in-memory cache for responses
- Async Support: Non-blocking operations for scalability
Feature Extensions
- Multi-Model Support: Switch between providers
- Fallback Models: Graceful degradation strategy
- Custom Tools: Domain-specific function calling
- Advanced Memory: Vector stores for long-term memory
- Evaluation: Automated testing and benchmarking
Code Quality
- Testing: Unit tests, integration tests
- Documentation: API docs, usage examples
- Type Safety: Stricter type hints
- Configuration: Externalized config files
Related Patterns
- Pattern 052
- Pattern 054
Pattern Families: – Reasoning Patterns (001-011) – Multi-Agent Patterns (012-019) – Tool Use Patterns (020-025) – Memory Patterns (026-031) – And many more…
Summary
The Mixture Of Agents pattern provides a powerful approach to building AI agents with specific capabilities. This implementation demonstrates:
ā Key Strengths: – Clean, maintainable code structure – LangChain integration for composability – Type hints and documentation – Production-ready patterns
šÆ Best Use Cases: – Applications requiring this pattern’s characteristics – Systems needing structured agent behavior – Production environments requiring reliability
š Learning Resources: – LangChain Documentation: https://python.langchain.com/ – Pattern Documentation: See related patterns – OpenAI API Reference: https://platform.openai.com/docs/
File: 053_mixture_of_agents.py
Total Lines: 656
Pattern Category: AI Agent Design Pattern #053
Generated from actual implementation code
For educational and reference purposes
Leave a Reply