GigaCode R&D has introduced GigaCode Universal Agent, an open-source Python project designed for agentic tasks. The system utilizes a core execution environment that connects LLMs via a chat/completions API to perform complex workflows like document analysis and code generation.
The architecture separates the agent's core from its specific skills and tools. Skills act as reusable protocols, while tools provide access to data and external systems via the Model Context Protocol. This modular design allows developers to add new scenarios without modifying the core agent cycle.
The system incorporates a 'judge' model that reviews the agent's execution trace and final output against the original task requirements. This ensures that tasks are not marked complete until all criteria, such as source verification and requirement fulfillment, are met.
Performance tests on the Qwen3.6-35B-A3B model showed that the Universal Agent outperformed OpenCode and Qwen Code on Verified and Terminal benchmarks. In cybersecurity tasks using DeepSeek-V4-Flash, adding a specific skill set increased the success rate from 70.6% to 76.7%.
The agent is currently being deployed within Sber for various internal tasks, including automated Pull Request generation and cybersecurity analysis. The project aims to provide a consistent execution environment that improves model performance across different domains.
