for the longest time, most people assumed the way to make ai better was simple:
build a bigger model, train on more data or add more parameters.
Microsoft Research just dropped a paper that challenges that idea and anyone following
@SentientAGI should check it out.
their system is called SkillOpt.
and what's interesting isn't the performance gains.
it's what they're optimizing.
instead of improving the model itself, they're improving the skill, the instructions ,the workflow and the accumulated experience the agent carries into a task.
every time the agent performs a task, SkillOpt analyzes the outcome, makes small edits to the skill, tests the change, and only keeps it if performance improves.
basically:
the model stays the same.
the skill gets smarter.
that distinction matters.
because we're entering a world where everyone has access to powerful models.
the intelligence is becoming accessible which means the advantage starts shifting elsewhere.
not who has the model but who has the best systems running on top of it.
the best workflows, the best accumulated knowledge, the best skills.
Microsoft tested SkillOpt across multiple benchmarks, models, and agent environments.
the optimized skills consistently outperformed the original versions.
but the bigger takeaway is what this says about the future.
we may spend less time training models and more time training the layer that sits above them.
because if the model is becoming a commodity, then the skill becomes the moat.
paper below 👇🏽
arxiv.org/abs/2605.23904