Minimax's version of hermesagent generate permanent skills from task completion
I started using MaxHermes a few weeks ago as a companion to Claude Code, mainly to see if the "self-evolving" pitch holds up in practice.
MaxHermes has a skill compounding feature, which is the agent builds procedural memory from actual task completions. Like after finishing a complex task, MaxHermes automatically evaluates what approach worked, and if certain conditions are met (like using multiple tools or receiving corrections), it generates a reusable skill document. Next time a similar task comes up, it calls that existing skill and runs faster and more accurately. The skills even get refined during use based on feedback.
I just think it's a really smart approach for it has optimized efficiency to the level of a "derivative" level.
I'm curious to hear from others who have experimented with self-evolving agents in production workflows. Has skill compounding actually changed how you work, or does it still feel like a novelty? For those who have used both MaxHermes and OpenClaw (the multi-agent orchestration tool MiniMax also offers), what's your take on which approach works better for what use cases?