On April 16, MiniMax (00100.HK) announced the launch of Hermes, the world's first cloud-based sandbox - MaxHermes. This assistant is built on the Hermes Agent, with its core breakthrough being the introduction of a "learning loop" mechanism, allowing it to autonomously extract reusable "Skills" from complex task execution and save them as independent documents. Unlike tools such as OpenClaw that rely on manually predefined capabilities, MaxHermes' skill library dynamically grows during use and self-iterates and improves based on user feedback.
In terms of technical implementation, MaxHermes integrates persistent cross-session memory, naturally language-defined scheduled tasks, and a parallel operation mechanism for multiple sub-agents, building an evolutionary intelligent agent capable of long-term operation. Currently, this assistant mainly uses MiniMax's latest released M2.7 programming model. This model has significantly improved in tool call accuracy, compliance with complex instructions, and Agent Harness compatibility, becoming one of the most active models in the Hermes ecosystem, providing a foundational computing power guarantee for the autonomous generation and precise execution of Skills.
The release of MaxHermes marks the transformation of AI Agents from "fixed capability tools" to "self-evolving entities." By integrating the two logics of learning loops and model iteration, MaxHermes greatly expands the capability boundary of AI in handling ambiguous and long-term tasks. This characteristic of continuously aligning with user preferences over usage time is expected to significantly enhance the practical penetration rate and delivery quality of AI in personal assistant and enterprise-level automation scenarios.
