ByteDance's MUSE-Autoskill builds agents that write their own skills
We spent the last few months watching humans hand-write skills for agents, the taste skills, the security skills, the design skills, all of it authored by people and dropped into a repo. MUSE-Autoskill, a new paper out of ByteDance, asks the obvious next question: why is the human still in that loop? Let the agent write its own skills.
The system is a full self-evolution loop with four moving parts. The agent creates skills when it runs into a task it can't cleanly handle. It stores them in memory. It manages that growing library, deciding what to keep, merge, or retire so it doesn't drown in its own past. And it evaluates whether a skill actually helped, feeding that signal back into what it creates next. The point is that none of these alone is enough. A skill creator with no manager becomes a hoarder; a manager with no evaluator can't tell good skills from clutter. MUSE wires all four together.
This lands squarely in a cluster that's been building all month, agents that improve themselves rather than waiting for the next model drop. We've seen MOSS rewrite its own source code and others edit their own prompts. MUSE works one layer up, at the skill level, which is arguably the sweet spot, more durable than a prompt tweak, less dangerous than rewriting your own code.
The honest caveat is this is a framework paper, not a product, and self-improving agents have a long history of looking great in a benchmark and brittle in the wild. But the direction is the one that matters. The endgame of the skills wave isn't a bigger marketplace of human-written skills. It's agents that grow their own and curate the collection without you.
Paper: arxiv.org/abs/2605.27366
← Back to all articles
The system is a full self-evolution loop with four moving parts. The agent creates skills when it runs into a task it can't cleanly handle. It stores them in memory. It manages that growing library, deciding what to keep, merge, or retire so it doesn't drown in its own past. And it evaluates whether a skill actually helped, feeding that signal back into what it creates next. The point is that none of these alone is enough. A skill creator with no manager becomes a hoarder; a manager with no evaluator can't tell good skills from clutter. MUSE wires all four together.
This lands squarely in a cluster that's been building all month, agents that improve themselves rather than waiting for the next model drop. We've seen MOSS rewrite its own source code and others edit their own prompts. MUSE works one layer up, at the skill level, which is arguably the sweet spot, more durable than a prompt tweak, less dangerous than rewriting your own code.
The honest caveat is this is a framework paper, not a product, and self-improving agents have a long history of looking great in a benchmark and brittle in the wild. But the direction is the one that matters. The endgame of the skills wave isn't a bigger marketplace of human-written skills. It's agents that grow their own and curate the collection without you.
Paper: arxiv.org/abs/2605.27366
Comments