gitwalter Agentic Loops
ReAct pattern implementation, reflection and self-correction, planning and task decomposition, iterative refinement patterns
Implement agentic reasoning loops - ReAct pattern, reflection, planning, and iterative refinement for autonomous agent behavior.
Process
- Review the task requirements.
- Apply the skill's methodology.
- Validate the output against the defined criteria.
Step 1: Basic ReAct Pattern
Implement the core ReAct loop: Reason → Act → Observe → Repeat:
from langchain_core.messages import HumanMessage, AIMessage, ToolMessage
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
from typing import List
llm = ChatOpenAI(model="gpt-4", temperature=0.7)
@tool
def search_knowledge_base(query: str) -> str:
"""Search internal knowledge base for information.
Args:
query: Search query
"""
# Implementation
return f"Knowledge base results for: {query}"
@tool
def calculate(expression: str) -> str:
"""Evaluate a mathematical expression.
Args:
expression: Math expression like '2 + 2'
"""
try:
result = eval(expression)
return str(result)
except Exception as e:
return f"Error: {str(e)}"
tools = [search_knowledge_base, calculate]
llm_with_tools = llm.bind_tools(tools)
async def react_loop(user_query: str, max_iterations: int = 10) -> str:
"""Basic ReAct pattern implementation."""
messages: List = [HumanMessage(content=user_query)]
iteration = 0
while iteration < max_iterations:
# REASON: LLM generates reasoning and decides on actions
response = await llm_with_tools.ainvoke(messages)
messages.append(response)
# Check if agent is done
if not response.tool_calls:
return response.content
# ACT: Execute tool calls
tool_map = {t.name: t for t in tools}
for tool_call in response.tool_calls:
tool_name = tool_call["name"]
tool_args = tool_call["args"]
tool_id = tool_call["id"]
if tool_name in tool_map:
try:
# OBSERVE: Get tool result
if asyncio.iscoroutinefunction(tool_map[tool_name].func):
result = await tool_map[tool_name].ainvoke(tool_args)
else:
result = tool_map[tool_name].invoke(tool_args)
messages.append(ToolMessage(
content=str(result),
tool_call_id=tool_id
))
except Exception as e:
messages.append(ToolMessage(
content=f"Error: {str(e)}",
tool_call_id=tool_id
))
iteration += 1
return "Maximum iterations reached."
# Usage
result = await react_loop("What is 15 * 23 and search for information about Python?")
Step 2: Reflection and Self-Correction
Add reflection step to evaluate and correct actions:
from langchain_core.prompts import ChatPromptTemplate
class ReflectiveAgent:
"""Agent with reflection and self-correction capabilities."""
def __init__(self, llm, tools):
self.llm = llm
self.tools = tools
self.llm_with_tools = llm.bind_tools(tools)
# Reflection prompt
self.reflection_prompt = ChatPromptTemplate.from_messages([
("system", """You are a reflective agent. After taking actions, evaluate:
1. Did the actions achieve the goal?
2. Were there any errors or issues?
3. What should be done differently?
4. Should we continue or stop?"""),
("user", "{history}")
])
async def reflect(self, messages: List, goal: str) -> dict:
"""Reflect on current progress toward goal."""
history = "\n".join([
f"{msg.__class__.__name__}: {msg.content}"
for msg in messages[-10:] # Last 10 messages
])
reflection = await self.reflection_prompt.ainvoke({
"history": f"Goal: {goal}\n\nConversation:\n{history}"
})
# Parse reflection
content = reflection.content.lower()
should_continue = "continue" in content or "retry" in content
errors_found = "error" in content or "failed" in content
return {
"reflection": reflection.content,
"should_continue": should_continue,
"errors_found": errors_found
}
async def execute(self, user_query: str, max_iterations: int = 10) -> str:
"""Execute with reflection."""
messages = [HumanMessage(content=user_query)]
iteration = 0
while iteration < max_iterations:
# Act
response = await self.llm_with_tools.ainvoke(messages)
messages.append(response)
if not response.tool_calls:
# Reflect before finishing
reflection = await self.reflect(messages, user_query)
if reflection["errors_found"] and reflection["should_continue"]:
# Self-correct
correction_msg = HumanMessage(
content=f"Reflection: {reflection['reflection']}\n\nPlease correct any errors."
)
messages.append(correction_msg)
continue
return response.content
# Execute tools
tool_map = {t.name: t for t in self.tools}
for tool_call in response.tool_calls:
tool_name = tool_call["name"]
tool_args = tool_call["args"]
tool_id = tool_call["id"]
if tool_name in tool_map:
try:
if asyncio.iscoroutinefunction(tool_map[tool_name].func):
result = await tool_map[tool_name].ainvoke(tool_args)
else:
result = tool_map[tool_name].invoke(tool_args)
messages.append(ToolMessage(
content=str(result),
tool_call_id=tool_id
))
except Exception as e:
messages.append(ToolMessage(
content=f"Error: {str(e)}",
tool_call_id=tool_id
))
# Reflect every few iterations
if iteration % 3 == 0:
reflection = await self.reflect(messages, user_query)
if not reflection["should_continue"]:
return f"Task completed. Reflection: {reflection['reflection']}"
iteration += 1
return "Maximum iterations reached."
# Usage
agent = ReflectiveAgent(llm, tools)
result = await agent.execute("Research Python best practices and summarize them")
Step 3: Planning and Task Decomposition
Break complex tasks into subtasks:
from pydantic import BaseModel, Field
from typing import List
class Subtask(BaseModel):
"""Represents a subtask in a plan."""
id: int
description: str
dependencies: List[int] = Field(default_factory=list)
status: str = "pending" # pending, in_progress, completed, failed
class Planner:
"""Agent that plans and decomposes tasks."""
def __init__(self, llm):
self.llm = llm
self.planning_prompt = ChatPromptTemplate.from_messages([
("system", """Break down the task into subtasks. For each subtask, identify:
1. What needs to be done
2. Dependencies on other subtasks
3. Order of execution
Return subtasks as a structured list."""),
("user", "{task}")
])
async def create_plan(self, task: str) -> List[Subtask]:
"""Decompose task into subtasks."""
response = await self.planning_prompt.ainvoke({"task": task})
# Parse response into subtasks (simplified - in practice use structured output)
# For now, return example structure
subtasks = [
Subtask(id=1, description="Research topic", dependencies=[]),
Subtask(id=2, description="Analyze findings", dependencies=[1]),
Subtask(id=3, description="Generate summary", dependencies=[2]),
]
return subtasks
async def execute_plan(self, task: str, tools: List) -> str:
"""Execute plan by running subtasks in order."""
plan = await self.create_plan(task)
results = {}
llm_with_tools = self.llm.bind_tools(tools)
for subtask in plan:
# Check dependencies
if any(results.get(dep_id) == "failed" for dep_id in subtask.dependencies):
subtask.status = "failed"
results[subtask.id] = "failed"
continue
subtask.status = "in_progress"
# Execute subtask
query = f"Subtask {subtask.id}: {subtask.description}"
messages = [HumanMessage(content=query)]
iteration = 0
while iteration < 5: # Max iterations per subtask
response = await llm_with_tools.ainvoke(messages)
messages.append(response)
if not response.tool_calls:
results[subtask.id] = response.content
subtask.status = "completed"
break
# Execute tools
tool_map = {t.name: t for t in tools}
for tool_call in response.tool_calls:
tool_name = tool_call["name"]
tool_args = tool_call["args"]
tool_id = tool_call["id"]
if tool_name in tool_map:
try:
if asyncio.iscoroutinefunction(tool_map[tool_name].func):
result = await tool_map[tool_name].ainvoke(tool_args)
else:
result = tool_map[tool_name].invoke(tool_args)
messages.append(ToolMessage(
content=str(result),
tool_call_id=tool_id
))
except Exception as e:
messages.append(ToolMessage(
content=f"Error: {str(e)}",
tool_call_id=tool_id
))
iteration += 1
if iteration >= 5:
subtask.status = "failed"
results[subtask.id] = "failed"
# Compile final result
return f"Plan execution complete. Results: {results}"
# Usage
planner = Planner(llm)
result = await planner.execute_plan("Research and summarize Python async patterns", tools)
Step 4: Iterative Refinement Pattern
Refine outputs through multiple iterations:
class RefinementAgent:
"""Agent that iteratively refines its output."""
def __init__(self, llm):
self.llm = llm
self.refinement_prompt = ChatPromptTemplate.from_messages([
("system", """You are an agent that improves outputs iteratively.
Given the current output and feedback, refine it to be better.
Continue refining until the output meets quality standards."""),
("user", """Original request: {request}
Current output: {current_output}
Feedback: {feedback}
Iteration: {iteration}
Refine the output based on the feedback.""")
])
async def refine(
self,
request: str,
initial_output: str = None,
max_iterations: int = 5,
quality_threshold: float = 0.8
) -> str:
"""Iteratively refine output."""
current_output = initial_output or ""
iteration = 0
while iteration < max_iterations:
# Generate or refine
if iteration == 0 and not initial_output:
# First iteration - generate initial output
response = await self.llm.ainvoke(request)
current_output = response.content
else:
# Refinement iteration
feedback = await self._evaluate_quality(request, current_output)
if feedback["score"] >= quality_threshold:
return current_output
response = await self.refinement_prompt.ainvoke({
"request": request,
"current_output": current_output,
"feedback": feedback["feedback"],
"iteration": iteration + 1
})
current_output = response.content
iteration += 1
return current_output
async def _evaluate_quality(self, request: str, output: str) -> dict:
"""Evaluate output quality (simplified - use LLM or metrics)."""
eval_prompt = ChatPromptTemplate.from_messages([
("system", """Evaluate the output quality on a scale of 0-1.
Provide a score and specific feedback for improvement."""),
("user", f"Request: {request}\n\nOutput: {output}")
])
response = await self.llm.ainvoke(eval_prompt.format(request=request, output=output))
# Parse score (simplified)
content = response.content.lower()
score = 0.7 # Default
if "excellent" in content or "perfect" in content:
score = 0.9
elif "good" in content:
score = 0.8
elif "needs improvement" in content:
score = 0.6
return {
"score": score,
"feedback": response.content
}
# Usage
refiner = RefinementAgent(llm)
refined = await refiner.refine(
"Write a comprehensive guide to Python async programming",
max_iterations=3
)
Step 5: Parallel Tool Execution
Execute multiple tools in parallel for efficiency:
import asyncio
from typing import List, Dict
async def parallel_react_loop(user_query: str, tools: List, max_iterations: int = 10) -> str:
"""ReAct loop with parallel tool execution."""
llm_with_tools = llm.bind_tools(tools)
messages = [HumanMessage(content=user_query)]
iteration = 0
while iteration < max_iterations:
response = await llm_with_tools.ainvoke(messages)
messages.append(response)
if not response.tool_calls:
return response.content
# Execute all tool calls in parallel
tool_map = {t.name: t for t in tools}
tool_tasks = []
for tool_call in response.tool_calls:
tool_name = tool_call["name"]
tool_args = tool_call["args"]
tool_id = tool_call["id"]
if tool_name in tool_map:
tool = tool_map[tool_name]
async def execute_tool(tool, args, tool_id):
try:
if asyncio.iscoroutinefunction(tool.func):
result = await tool.ainvoke(args)
else:
result = tool.invoke(args)
return ToolMessage(
content=str(result),
tool_call_id=tool_id
)
except Exception as e:
return ToolMessage(
content=f"Error: {str(e)}",
tool_call_id=tool_id
)
tool_tasks.append(execute_tool(tool, tool_args, tool_id))
# Wait for all tools to complete
tool_results = await asyncio.gather(*tool_tasks)
messages.extend(tool_results)
iteration += 1
return "Maximum iterations reached."
Step 6: Conditional Loop Control
Add conditions to control loop behavior:
class ConditionalAgent:
"""Agent with conditional loop control."""
def __init__(self, llm, tools):
self.llm = llm
self.tools = tools
self.llm_with_tools = llm.bind_tools(tools)
async def execute(
self,
user_query: str,
stop_conditions: List[str] = None,
max_iterations: int = 10,
min_confidence: float = 0.7
) -> str:
"""Execute with conditional stopping."""
messages = [HumanMessage(content=user_query)]
iteration = 0
stop_conditions = stop_conditions or []
while iteration < max_iterations:
response = await self.llm_with_tools.ainvoke(messages)
messages.append(response)
# Check stop conditions
content_lower = response.content.lower()
if any(condition.lower() in content_lower for condition in stop_conditions):
return f"Stopped due to condition. {response.content}"
if not response.tool_calls:
# Check confidence (simplified)
confidence = await self._estimate_confidence(response.content)
if confidence >= min_confidence:
return response.content
else:
# Low confidence - try again
messages.append(HumanMessage(
content="The response seems uncertain. Please try again with more confidence."
))
continue
# Execute tools
tool_map = {t.name: t for t in self.tools}
for tool_call in response.tool_calls:
tool_name = tool_call["name"]
tool_args = tool_call["args"]
tool_id = tool_call["id"]
if tool_name in tool_map:
try:
if asyncio.iscoroutinefunction(tool_map[tool_name].func):
result = await tool_map[tool_name].ainvoke(tool_args)
else:
result = tool_map[tool_name].invoke(tool_args)
messages.append(ToolMessage(
content=str(result),
tool_call_id=tool_id
))
except Exception as e:
messages.append(ToolMessage(
content=f"Error: {str(e)}",
tool_call_id=tool_id
))
iteration += 1
return "Maximum iterations reached."
async def _estimate_confidence(self, content: str) -> float:
"""Estimate confidence in response (simplified)."""
uncertainty_words = ["maybe", "perhaps", "uncertain", "not sure"]
content_lower = content.lower()
if any(word in content_lower for word in uncertainty_words):
return 0.5
return 0.8
# Usage
agent = ConditionalAgent(llm, tools)
result = await agent.execute(
"Research Python async patterns",
stop_conditions=["error", "cannot"],
min_confidence=0.8
)
### Step 7: Sequential Thinking Integration (New)
Use the `sequential-thinking` MCP server for complex problem solving:
```python
# Decompose a complex problem
response = await client.chat.completions.create(
messages=[{"role": "user", "content": "Design a scalable microservices architecture for a banking app"}],
tools=[{
"type": "mcp",
"name": "sequential-thinking",
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
}]
)
Agentic Loop Patterns
| Pattern | Description | Use Case | ||-|-| | ReAct | Reason → Act → Observe → Repeat | General agentic tasks | | Reflection | Act → Reflect → Correct → Repeat | Quality-critical tasks | | Planning | Plan → Decompose → Execute → Aggregate | Complex multi-step tasks | | Refinement | Generate → Evaluate → Refine → Repeat | Content generation | | Parallel | Act → Execute tools in parallel → Observe | Performance-critical | | Conditional | Act → Check conditions → Continue/Stop | Controlled execution |
Best Practices
- Set appropriate
max_iterationsto prevent infinite loops - Implement proper error handling for tool calls
- Use reflection for quality-critical tasks
- Decompose complex tasks into subtasks
- Execute independent tools in parallel
- Add stop conditions for controlled execution
- Monitor token usage in long-running loops
- Log loop iterations for debugging
- Use structured outputs for planning
- Implement confidence thresholds
Anti-Patterns
| Anti-Pattern | Fix |
|---|---|
| No iteration limit | Set max_iterations |
| Sequential tool execution | Use asyncio.gather for parallel execution |
| No error handling | Wrap tool calls in try/except |
| No reflection | Add reflection step for quality |
| Ignoring tool errors | Report errors back to agent |
| No planning | Decompose complex tasks |
| Infinite loops | Add stop conditions |
| No confidence checking | Evaluate response quality |
| Synchronous execution | Use async/await throughout |
| No logging | Log iterations and decisions |
Related
- Knowledge:
{directories.knowledge}/agentic-loop-patterns.json - Skill:
anthropic-patterns - Skill:
tool-usage - Skill:
using-langchain - Skill:
langgraph-agent-building
When to Use
This skill should be used when strict adherence to the defined process is required.
Prerequisites
- Basic understanding of the agent factory context.
- Access to the necessary tools and resources.