SKILL-DISCO: agents that compile their own shortcuts

Here's a clean idea from a new paper out of Microsoft Research Asia (arXiv 2606.26669). Agents waste an enormous amount of effort re-solving the same kind of task from scratch every single time. SKILL-DISCO watches an agent's successful runs, treats each one as a path through a state-transition graph, pulls out the sub-paths that keep recurring, and compiles them into callable, parameterized, executable procedural skills. Not a vague remember-this note, but actual FSM-style routines the agent can invoke like functions.

On ALFWorld and WebArena it lifted success rates while cutting the number of turns, and it held up across model sizes. Worth a note for one reason: most of the skill-learning work lately distills skills into the weights, LatentSkill, OPID just two days ago. This goes the other way. Skills as compiled, inspectable control flow that sits outside the model.

That's the more debuggable path, and arguably the more durable one, because you can actually read what the agent learned and fix it when it's wrong. The recurring thesis of the quarter, stated one more time: the moat isn't the model, it's the accumulated procedure. Whether it lives in weights or in compiled skill files is now the live argument, and SKILL-DISCO makes the strongest recent case for files. Link: arxiv.org/abs/2606.26669