Open-source DeepResearch enables an open agent framework for LLMs with code actions and web browsing

AI Impact Summary

An open-source initiative to reproduce and share the agent framework behind DeepResearch, enabling LLMs to perform tool-based actions (web browsing, code execution) via a code-first action layer. The team cites GAIA validation scores reaching 55.15% with code-based agents, a significant improvement over JSON-only baselines, indicating practical gains for multi-step, tool-using tasks. This provides technical teams with a ready-to-experiment stack (smolagents, Magentic-One tooling) to deploy agent-driven LLM workflows, potentially reducing vendor lock-in and accelerating iteration, while increasing governance, security, and data-handling considerations for web-enabled agents.