Microsoft SkillOpt boosts agent performance with self-evolving skills

SkillOpt is an important milestone in turning agents from hand-tuned prompt soups into objects of disciplined optimization. Instead of fine-tuning model weights, Microsoft treats the skill document that governs an agent’s behavior as “external state” and subjects it to a loop that looks a lot like gradient descent: rollout tasks, reflect on success and failure, propose bounded edits, and accept only if held-out performance improves. The fact that this works across GPT‑5.x and Qwen models, and across Codex- and Claude-based agent harnesses, suggests this is a generally useful recipe, not a one-off trick.

For the AGI race, this effectively creates a new axis of improvement orthogonal to raw model scale. If you can take a frozen model and boost its performance 20–30 points on hard benchmarks just by systematically training its procedures, you extend the useful life of each frontier model generation and make smaller models more competitive. That favors players with strong tooling and evaluation pipelines, not just those with the largest GPUs.

It also moves us closer to agents that can iteratively refine their own operating manuals under constraints, a capability that will matter for long-lived, safety‑critical systems. The big open question is how to align the objectives of these skill optimizers with human intent when they’re deployed in messy real-world workflows rather than clean benchmarks.