cogency 0.9.3

Smart AI agents that think as hard as they need to.

Cogency

Smart AI agents that think as hard as they need to.

🚧 Production Beta (v0.9.1) - Architecturally complete, actively gathering feedback from early adopters. Ready for serious evaluation and beta deployments.

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

# Simple task → thinks fast
await agent.run("What's 2+2?")

# Complex task → thinks deep
await agent.run("Analyze this codebase and suggest architectural improvements")
# Automatically escalates reasoning, uses relevant tools, streams thinking

🧠 Adaptive Reasoning

Thinks fast or deep as needed - agents discover task complexity during execution and adapt their cognitive approach automatically.

• Fast React: Direct execution for simple queries
• Deep React: Reflection + planning for complex analysis
• Zero-cost switching: Seamless transitions preserve full context
• Runtime discovery: No upfront classification - intelligence governs intelligence

🚀 Key Features

• 🤖 Agents in 3 lines - Fully functional, tool-using agents from a single import
• 🔥 Adaptive reasoning - Thinks fast or deep as needed, switches seamlessly during execution
• 🌊 Streaming first - Watch agents think in real-time with full transparency
• 🛠️ Automatic tool discovery - Drop tools in, they auto-register and route intelligently
• ⚡️ Zero configuration - Auto-detects LLMs, tools, memory from environment (fast models by default)
• 🧠 Built-in memory - Persistent memory with extensible backends (Pinecone, ChromaDB, PGVector)
• ✨ Clean tracing - Every reasoning step traced and streamed with clear phase indicators
• 🌍 Universal LLM support - OpenAI, Anthropic, Gemini, Grok, Mistral out of the box
• 🧩 Extensible design - Add tools, memory backends, embedders with zero friction
• 👥 Multi-tenancy - Built-in user contexts and conversation isolation
• 🏗️ Production hardened - Resilience, rate limiting, metrics, tracing included

How It Works

Prepare → Reason → Act → Respond

👤 Plan a Tokyo trip with $2000 budget

🔧 Tools: code, search
🧠 Task complexity → escalating to Deep React
🌤️ weather(Tokyo) → 25°C sunny, rain Thu-Fri
🧮 code(print(2000 / 5)) → 400.0
🔍 search(Tokyo indoor activities) → Museums, temples
💭 Reflection: Need indoor backup plans for rainy days
📋 Planning: 5-day itinerary with weather contingencies
🤖 Here's your optimized Tokyo itinerary...

The preparing phase handles tool selection, memory operations, and intelligent routing between reasoning modes.

Quick Examples

Custom Tools

from cogency.tools import Tool, tool

@tool
class MyTool(Tool):
    def __init__(self):
        super().__init__("my_tool", "Does something useful")

    async def run(self, param: str):
        return {"result": f"Processed: {param}"}

# Tool auto-registers - just create agent
agent = Agent("assistant")
await agent.run("Use my_tool with hello")

Real-World Applications

# Research Agent
agent = Agent("researcher")
await agent.run("Latest quantum computing developments?")

# Coding Assistant
agent = Agent("coder")
await agent.run("Fix the auth bug in this Flask app")

# Data Analyst
agent = Agent("analyst")
await agent.run("Analyze sales trends in quarterly_data.csv")

Built-in Tools

Agents automatically discover and use relevant tools:

💻 Code - Execute Python code
🔍 Search - Web search for current information
🌤️ Weather - Current conditions and forecasts
📁 Files - Create, read, edit, list, delete files
💻 Shell - Execute system commands safely
🐍 Code - Python code execution in sandboxed environment
📊 CSV - Data processing and analysis
🗄️ SQL - Database querying and management
🌐 HTTP - Make HTTP requests with JSON parsing
🕒 Time - Date/time operations and timezone conversions
🔗 Scrape - Web scraping with content extraction
🧠 Recall - Memory search and retrieval

Installation

pip install cogency

Beta Note: Cross-provider testing is ongoing. OpenAI and Anthropic are well-tested; other providers may have edge cases.

Set any LLM API key:

export OPENAI_API_KEY=...     # or
export ANTHROPIC_API_KEY=...  # or
export GEMINI_API_KEY=...        # etc

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 backends and configuration
• Reasoning - Adaptive reasoning modes
• Configuration - Advanced configuration options
• Troubleshooting - Common issues and solutions

📄 License

Apache 2.0 - Build whatever you want.

Beta Feedback

We're actively gathering feedback from early adopters:

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Known limitations: Cross-provider behavior, memory backend edge cases

---

Built for developers who want agents that just work, not frameworks that require PhD-level configuration.