MiniMax MaxHermes: the agent that writes its own skill library

MiniMax just shipped MaxHermes, a cloud-sandboxed AI agent that turns every task you throw at it into a reusable skill. This is the first productized version of their Hermes Agent research, and the pitch is sharp: instead of pre-defining tools like OpenClaw does, MaxHermes watches you work, extracts the playbook, and saves it as a standalone skill document that it reuses next time.

The learning loop is the whole point. Every long-running task gets compressed into a named skill. The skill library grows as you use the product. Over time, what used to take a prompt-engineering session becomes a one-click invocation. MiniMax calls this "one-click horse-raising mode" — feed it tasks, watch the agent get smarter without retraining the model.

Under the hood it runs on M2.7, MiniMax's latest coding-oriented model. Persistent cross-session memory, natural-language scheduled tasks, parallel sub-agents — the full autonomous-work stack. And critically for enterprise, it ships with Feishu, DingTalk, and WeCom integrations out of the box, so no sysadmin has to provision servers or wire up API keys. Token-based pricing replaces the usual per-seat SaaS math.

This is MiniMax's answer to Claude Managed Agents and Codex — the China version of cloud-native autonomous work. The bet: skills that compound over months are worth more than a bigger context window.

Try it at https://agent.minimax.io/max-hermes