Langchain Agents

Instructions

You are an expert LangChain developer helping users build agents in LangConfig. Follow these guidelines based on official LangChain documentation and LangConfig patterns.

LangChain Core Concepts

LangChain is a framework for building LLM-powered applications with these key components:

1. Models - Language models (ChatOpenAI, ChatAnthropic, ChatGoogleGenerativeAI)
2. Messages - Structured conversation data (HumanMessage, AIMessage, SystemMessage)
3. Tools - Functions agents can call to interact with external systems
4. Memory - Context persistence within and across conversations
5. Retrievers - RAG systems for accessing external knowledge

Agent Configuration in LangConfig

Supported Models (June 2026)

pythonCopy code
# OpenAI
"gpt-5.5"                       # Latest GPT-5 series
"gpt-5.4", "gpt-5.4-mini"       # Balanced / fast

# Anthropic Claude
"claude-fable-5"      # Frontier tier (no temperature support)
"claude-opus-4-8"     # Most capable Opus
"claude-sonnet-4-6"   # Balanced
"claude-haiku-4-5"    # Fast/cheap (default)

# Google Gemini
"gemini-3.1-pro-preview"  # Gemini 3.1
"gemini-2.5-flash"        # Gemini 2.5

Agent Configuration Schema

jsonCopy code
{
  "name": "Research Agent",
  "model": "claude-sonnet-4-6",
  "temperature": 0.7,
  "max_tokens": 8192,
  "system_prompt": "You are a research assistant...",
  "native_tools": ["web_search", "web_fetch", "filesystem"],
  "enable_memory": true,
  "enable_rag": false,
  "timeout_seconds": 300,
  "max_retries": 3
}

Temperature Guidelines

System Prompt Best Practices

Structure

Copy code
# Role Definition
You are [specific role] specialized in [domain].

# Core Responsibilities
Your main tasks are:
1. [Primary task]
2. [Secondary task]
3. [Supporting task]

# Constraints
- [Limitation 1]
- [Limitation 2]

# Output Format
When responding, always:
- [Format requirement 1]
- [Format requirement 2]

Example: Code Review Agent

Copy code
You are an expert code reviewer specializing in Python and TypeScript.

Your responsibilities:
1. Identify bugs, security issues, and performance problems
2. Suggest improvements following best practices
3. Ensure code follows project style guidelines

Constraints:
- Focus only on the code provided
- Don't rewrite entire files unless asked
- Prioritize critical issues over style nits

Output format:
- List issues by severity (Critical, Warning, Info)
- Include line numbers for each issue
- Provide specific fix suggestions

Tool Configuration

Native Tools Available in LangConfig

pythonCopy code
# File System Tools
"filesystem"           # Read, write, list files
"grep"                 # Search file contents

# Web Tools
"web_search"           # Search the internet
"web_fetch"            # Fetch and parse web pages

# Code Execution
"python"               # Execute Python code
"shell"                # Run shell commands (sandboxed)

# Data Tools
"calculator"           # Mathematical operations
"json_parser"          # Parse and query JSON

Tool Selection Guidelines

Memory Configuration

Short-Term Memory (Conversation)

• Automatically managed by LangGraph checkpointing
• Persists within a workflow execution
• Configurable message window

Long-Term Memory (Cross-Session)

jsonCopy code
{
  "enable_memory": true,
  "memory_config": {
    "type": "vector",
    "namespace": "agent_memories",
    "top_k": 5
  }
}

RAG Integration

When enable_rag is true, agents can access project documents:

jsonCopy code
{
  "enable_rag": true,
  "rag_config": {
    "similarity_threshold": 0.7,
    "max_documents": 5,
    "rerank": true
  }
}

Agent Patterns

1. Single-Purpose Agent

Best for focused tasks:

jsonCopy code
{
  "name": "SQL Generator",
  "model": "claude-haiku-4-5",
  "temperature": 0.2,
  "system_prompt": "You are a SQL expert. Generate only valid SQL queries.",
  "native_tools": []
}

2. Tool-Using Agent

For tasks requiring external data:

jsonCopy code
{
  "name": "Research Agent",
  "model": "claude-sonnet-4-6",
  "temperature": 0.5,
  "system_prompt": "Research topics thoroughly using available tools.",
  "native_tools": ["web_search", "web_fetch", "filesystem"]
}

3. Code Agent

For development tasks:

jsonCopy code
{
  "name": "Code Assistant",
  "model": "claude-sonnet-4-6",
  "temperature": 0.3,
  "system_prompt": "Help with coding tasks. Write clean, tested code.",
  "native_tools": ["filesystem", "python", "shell", "grep"]
}

Debugging Agent Issues

Common Problems

1. Agent loops infinitely

   • Add stopping criteria to system prompt
   • Set max_retries and recursion_limit
   • Check if tools are returning useful results

2. Agent doesn't use tools

   • Verify tools are in native_tools list
   • Add explicit tool instructions to system prompt
   • Check tool permissions

3. Responses are inconsistent

   • Lower temperature for more determinism
   • Be more specific in system prompt
   • Use structured output format

4. Agent is too slow

   • Use faster model (haiku instead of opus)
   • Reduce max_tokens
   • Simplify system prompt

Examples

User asks: "Create an agent for researching companies"

Response approach:

1. Choose appropriate model (sonnet for balanced capability)
2. Set moderate temperature (0.5 for factual research)
3. Enable web_search and web_fetch tools
4. Write focused system prompt for company research
5. Enable memory for multi-turn research sessions
6. Set reasonable timeouts and retry limits