Awesome AI Agents › AI Agent Gateway

Upsonic/Upsonic

⭐ 7954 Python repository created 2024-05-26

Upsonic is a highly reliable AI agent framework designed to support real-world applications with a focus on trustworthiness and accuracy. It addresses critical challenges in AI agent deployment by providing advanced reliability features such as verification layers, triangular architecture, validator agents, and output evaluation systems. The framework emphasizes production readiness by enabling easy activation of reliability layers without disrupting existing functionalities. A key innovation is the Model Context Protocol (MCP), which allows seamless integration and use of various tools developed both officially and by third parties, eliminating the need for custom tool development from scratch. Upsonic supports integrated browser and computer use, enabling agents to operate on non-API systems, and offers a secure runtime environment with isolated execution for enhanced security. The reliability layer is a standout feature, incorporating verifier agents that validate outputs for inconsistencies, numerical errors, and hallucinations, and editor agents that refine outputs based on verifier feedback. The system uses iterative verification rounds and controlled feedback loops to ensure high output accuracy. Benchmark tests demonstrate Upsonic's superior reliability compared to other frameworks, achieving a 99.3% reliability score in transforming JSON keys without hard-coded changes. Upsonic is designed for scalability and can be deployed on cloud platforms like AWS and GCP or locally via Docker. It supports task-centric design with options for basic tasks using LLM calls, advanced tasks with V1 agents, and complex automation with V2 agents integrated with MCP. The framework includes a MCP server for multi-client processing, a tool-calling server for secure tool management, and integration with Anthropic's 'Computer Use' for human-like task execution. Adding custom tools or MCP tools is straightforward, requiring minimal code. Comprehensive documentation is available online, and the framework supports Python 3.10+, with compatibility for OpenAI, Anthropic, Azure, and Bedrock APIs. Example code demonstrates agent creation, task execution, tool integration via MCP, and multi-agent task distribution, highlighting Upsonic's flexibility and power in building reliable AI agent workflows.

https://github.com/Upsonic/Upsonic

agentagent-frameworkai-agent-frameworkanthropicawsclaudecomputer-usedockerfeedback-loopsgcpisolated-environmentiterative-verificationllmsmcpmodel-context-protocolmulti-client-processingopenaioutput-evaluationproduction-readypythonragreliabilityscalabilitysecure-runtimetask-centric-designtool-integrationvalidator-agentsverification-layers