cogency 1.2.0

A reasoning engine for adaptive AI agents

Cogency

A reasoning engine for adaptive AI agents.

from cogency import Agent
agent = Agent("assistant")

# Simple task → direct response
agent.run("What's 2+2?")

# Complex task → adaptive reasoning
agent.run("Analyze this codebase and suggest architectural improvements")
# Automatically escalates reasoning depth and tool usage

Why Cogency?

Zero ceremony, maximum capability - get production-ready agents from a single import.

• 🔒 Semantic security - Built-in safety, blocks unsafe requests automatically
• ⚡ Adaptive reasoning - Thinks fast for simple queries, deep for complex tasks
• 🛠️ Smart tooling - Tools auto-register and route intelligently
• 🧠 Built-in memory - Persistent context that actually learns about users
• 🏗️ Production ready - Resilience, tracing, and error recovery out of the box

Get Started in 30 Seconds

pip install cogency
export OPENAI_API_KEY=...

from cogency import Agent

agent = Agent("assistant")
result = agent.run("What's in the current directory?")
print(result)

That's it. No configuration, no setup, no tool registration. Just working agents.

What Makes It Different

Semantic Security

agent.run("rm -rf /")  # ❌ Blocked automatically
agent.run("List files safely")  # ✅ Proceeds normally

Adaptive Intelligence

agent.run("What's 2+2?")  # Fast: Direct response
agent.run("Analyze my codebase")  # Deep: Multi-step reasoning

Memory That Actually Works

agent = Agent("assistant", memory=True)
agent.run("I prefer Python and work at Google")
agent.run("What language should I use?")  # → "Python"

Built-in Capabilities

Tools that just work: 📁 Files - Read, write, edit any file
💻 Shell - Execute commands safely
🌐 HTTP - API calls and requests
📖 Scrape - Extract web content
🔍 Search - Web search via DuckDuckGo

Plus add your own:

@tool
class DatabaseTool(Tool):
    async def run(self, query: str):
        return await db.execute(query)

# Automatically available to all agents

Universal LLM Support

Works with any LLM - just set the API key:

# OpenAI
export OPENAI_API_KEY=sk-...

# Anthropic  
export ANTHROPIC_API_KEY=sk-ant-...

# Google Gemini
export GEMINI_API_KEY=...

# Mistral
export MISTRAL_API_KEY=...

# OpenRouter (cost-effective)
export OPENROUTER_API_KEY=sk-or-v1-...

# Groq (high-performance) 
export GROQ_API_KEY=gsk_...

# Ollama (local models)
export OLLAMA_API_KEY=...

No configuration needed - Cogency detects and configures automatically.

Production Features

Streaming responses:

async for chunk in agent.stream("Analyze this large codebase"):
    print(chunk, end="")

Full observability:

result = agent.run("Deploy my app")
logs = agent.logs()  # See exactly what happened
print(logs)  # ["🔧 triage: selected 2 tools", "💻 shell: deploying...", ...]

Error resilience:

# Tool failures don't crash execution
agent.run("List files in /nonexistent")  # → Graceful error handling
# API timeouts auto-retry with backoff
# Memory failures don't block responses

Advanced Usage

# Full customization when needed
agent = Agent(
    "assistant",
    memory=True,              # Persistent user context
    tools=["files", "shell"],  # Specific tools only  
    max_iterations=20,        # Deep reasoning limit
    debug=True               # Detailed execution logs
)

# Custom memory configuration
from cogency.config import MemoryConfig
agent = Agent("assistant", memory=MemoryConfig(threshold=8000))

# Custom event handlers
agent = Agent("assistant", handlers=[websocket_handler])

Documentation

• Quick Start - Get running in 5 minutes
• API Reference - Complete Agent class documentation
• Tools - Built-in tools and custom tool creation
• Examples - Detailed code examples and walkthroughs
• Memory - Memory system documentation
• Reasoning - Adaptive reasoning modes

License

Apache 2.0

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions

Built for developers who want agents that just work.