The model is still not the product (adlrocha.substack.com)

A recent analysis highlights that in AI agent development, the true innovation lies not within the models themselves but in the engineering that surrounds them. The post reflects on the Claude Code leak and the patterns across various open-source projects like Hermes, Pi, and Opencode, emphasizing how these platforms address crucial challenges in agent functionality—namely, context management, memory structure, and composable tools. The author proposes that there is a convergence toward best practices among developers tackling similar problems in the realm of large language models (LLMs). Focusing on Hermes, which has gained significant traction for its unique architecture, the author notes the project's modular design comprising six layers that facilitate clear interactions among components. A standout feature is Hermes’ self-authoring skills system, allowing agents to dynamically create, maintain, and refine skills based on past experiences. This adaptability is pivotal for evolving agent capabilities and enhancing their performance over time. Additionally, the context compression mechanism in Hermes ensures that agents maintain efficient memory use during lengthy interactions, a critical capability for sustained operation. Overall, these innovations not only improve agent performance but also set the stage for more adaptive and efficient software engineering practices in AI and ML.