cogency 0.4.0

Agentic AI out of the box

Cogency

3-line AI agents that just work

Cogency is a multistep reasoning framework that makes building AI agents stupidly simple. Auto-detects providers, intelligently routes tools, streams transparent reasoning.

Installation

# Current broken until next release v0.3.1 (pending)
# Recommended: Fork and install from source
pip install cogency

Quick Start

import asyncio
from cogency import Agent

async def main():
    # That's it. Auto-detects LLM from .env
    agent = Agent("assistant")
    result = await agent.run("What is 25 * 43?", mode="summary")
    print(result)

asyncio.run(main())

Requirements: Python 3.9+, API key in .env

Core Philosophy

• Zero ceremony - 3 lines to working agent
• Magical auto-detection - Detects OpenAI, Anthropic, Gemini, Grok, Mistral from environment
• Intelligent tool routing - LLM filters relevant tools, keeps prompts lean
• Stream-first - Watch agents think in real-time
• Plug-and-play - Drop in new tools, they just work

Examples

Hello World

import asyncio
from cogency import Agent

async def main():
    agent = Agent("assistant")
    result = await agent.run("What is 25 * 43?", mode="summary")
    print(result)  # 1075

asyncio.run(main())

With Tools

import asyncio
from cogency import Agent, WeatherTool

async def main():
    agent = Agent("weather_assistant", tools=[WeatherTool()])
    result = await agent.run("What's the weather in San Francisco?", mode="summary")
    print(result)

asyncio.run(main())

Streaming Reasoning

import asyncio
from cogency import Agent, CalculatorTool, WebSearchTool

async def main():
    agent = Agent("analyst", tools=[CalculatorTool(), WebSearchTool()])

    # Stream with trace mode for full visibility
    async for chunk in agent.stream("Find Bitcoin price and calculate value of 0.5 BTC", mode="trace"):
        print(chunk, end="", flush=True)

asyncio.run(main())

Custom Tools

from cogency import Agent, BaseTool

class TimezoneTool(BaseTool):
    def __init__(self):
        super().__init__("timezone", "Get time in any city")
    
    async def run(self, city: str):
        return {"time": f"Current time in {city}: 14:30 PST"}
    
    def get_schema(self):
        return "timezone(city='string')"

agent = Agent("time_assistant", tools=[TimezoneTool()])

Installation

pip install cogency

# Set your API keys
echo "OPENAI_API_KEY=sk-..." >> .env
echo "ANTHROPIC_API_KEY=sk-ant-..." >> .env
echo "GEMINI_API_KEY=your-key-here" >> .env
# Agent auto-detects from any available key

Supported Providers

LLMs (Auto-detected):

• OpenAI (GPT-4, GPT-3.5)
• Anthropic (Claude)
• Google (Gemini)
• xAI (Grok)
• Mistral

Embeddings (Auto-detected):

• OpenAI (text-embedding-3)
• Nomic (nomic-embed-text)
• Sentence Transformers (local)

Built-in Tools:

• Calculator - Math operations
• Weather - Real weather data (no API key)
• Timezone - World time (no API key)
• WebSearch - Internet search
• FileManager - File operations

PRARR Architecture

Cogency uses Plan-Reason-Act-Reflect-Respond for transparent multi-step reasoning:

📋 PLAN    → Analyze request, filter relevant tools
🧠 REASON  → Determine tool usage strategy  
⚡ ACT     → Execute tools, gather results
🔍 REFLECT → Filter out bullshit, focus results
💬 RESPOND → Generate clean final answer

Every step is streamable and traceable.

Output Examples

Output Modes

Cogency supports three output modes:

Summary Mode (Default): Clean final answer only

result = await agent.run("What's 15 * 23?", mode="summary")
print(result)  # "345"

Trace Mode: Beautiful execution trace + answer

result = await agent.run("What's 15 * 23?", mode="trace")
# Outputs:
# 🚀 EXECUTION TRACE (450ms total)
# ==================================================
# 🔸 PLAN [14:30:15] 120ms
#    📥 'What's 15 * 23?'
#    📤 Decision: tool_needed
#
# 🔸 ACT [14:30:15] 200ms
#    📥 calculator(expression="15 * 23")
#    📤 Result: 345
# ==================================================
# ✅ Final: 345

Dev Mode: Raw state dumps for debugging

result = await agent.run("What's 15 * 23?", mode="dev")
# Outputs full AgentState with all internal data

Streaming: Real-time output with any mode

async for chunk in agent.stream("Calculate something", mode="trace"):
    print(chunk, end="", flush=True)

Key Features

• Auto-detection - Zero config provider setup
• Tool subsetting - Intelligent filtering keeps prompts lean
• Key rotation - Load balance across multiple API keys
• Result filtering - Remove execution metadata in REFLECT
• Stream transparency - Watch reasoning in real-time
• Beautiful traces - Human-readable execution logs
• Plug-and-play - Drop in tools, they auto-register

Contributing

Framework designed for extension:

# Add new LLM provider
class YourLLM(ProviderMixin, BaseLLM):
    async def invoke(self, messages, **kwargs):
        # Your implementation
        
# Add new tool  
class YourTool(BaseTool):
    async def run(self, **params):
        # Your implementation

That's it. Auto-discovery handles the rest.

License

MIT - Build whatever you want.

---

Cogency: AI agents without the ceremony.