Microsoft SkillOpt upgrades AI agent skills without changing model weights

AI agents are finally getting the “real work” everyone promised, but there’s a snag: their behavior is often controlled by skills that live in folders of markdown (.md) files, not by model weights. Those skills are instructions, tool-use policies, output constraints, and known failure modes inserted into an agent’s context before execution. And while that design lets teams customize an underlying model without touching its parameters, it also creates a slow, error-prone upgrade loop. If you want better performance, you typically retype and revise every file by hand, trying edits like a guessing game.

SkillOpt, a new open-source (MIT Licensed) Microsoft framework, attacks that exact bottleneck by optimizing the skill document itself, not the model. It takes a skill.md file and treats it like a trainable object that evolves based on performance feedback, using deep-learning-style optimization controls. Most importantly, it does all that without making changes to the underlying model’s weights, which is the part enterprises usually can’t afford to touch. In evaluations, SkillOpt outperforms baselines across 52 combinations of model, benchmark, and harness, including models like GPT-5.5 and Qwen.

Here’s why this matters beyond “cool research.” Agent skill files are supposed to give companies procedural reliability: consistent formats, tool policies, and guardrails that reduce embarrassing failures. But the moment you automate revisions to those text artifacts, you run into a stability problem that deep learning doesn’t let you ignore. As Yifan Yang, Senior Research SDE at Microsoft Research Asia, told VentureBeat, the breaking point isn’t whether a team can change a skill. It’s that they can’t guarantee the change is an improvement. He points to three recurring failure modes: no step-size control, so skills drift; no validation, so plausible edits silently regress performance; and no negative memory, so the same failed edit keeps coming back.

Yang also gives a concrete example of how ugly “ungated” rewrites can get. An ungated rewrite pushed GPT-5.5 on SpreadsheetBench from 41.8 down to 41.1. That’s the hidden cost executives should care about: when revisions are unstructured, performance can fall even though the text still looks reasonable. And those risks get amplified in multi-step workflows, where frontier models are weakest in procedural discipline rather than pure reasoning. In other words, the failure often isn’t “the model can’t think,” it’s “the agent can’t follow the playbook reliably across steps.” If your enterprise deployment depends on tool policy and format adherence, a regression is not a rounding error. It’s an outage with better typography.

SkillOpt’s core idea is to import mathematical discipline into text optimization by using an iterative propose-and-test loop. It separates the model executing tasks from the model optimizing the skill. The target model runs batches of tasks to generate execution trajectories, which become evidence for the current skill state. Then an offline optimizer model analyzes those trajectories to separate successes from failures into minibatches, so it can identify systematic procedural errors rather than chasing one-off anomalies. Based on these patterns, it proposes edits to the skill document: structural add, delete, or replace operations.

But SkillOpt does not treat every suggestion as gospel. Proposed edits are reviewed to filter duplicates or contradictions, then the optimizer ranks candidate edits by expected utility. Next comes the stability mechanic executives should recognize, even if they never learned the math: SkillOpt clips the edit list to a maximum edit budget for that step, effectively limiting how far each update can move away from the prior skill. The candidate skill is then evaluated on a held-out validation set using the target model. If it improves validation score, the edit set is accepted and becomes the current skill. If it fails, those edits are rejected into a rejected-edit buffer, providing negative feedback so the optimizer avoids repeating the same mistake. SkillOpt also describes the deep-learning analogy as operational rather than decorative, and it uses epoch-level updates that compare tasks under previous versus current skills, acting like a momentum term to carry durable procedural lessons forward while isolating fast step-level changes.

So what does “better” look like in practice? Researchers tested SkillOpt across frontier models like GPT-5.5 and smaller models including GPT-5.4-mini and Qwen3.5-4B, and they deployed skills inside different execution harnesses, from plain chat to complex coding harnesses like the Codex CLI and Claude Code. Evaluations covered single-round question answering, multi-round code generation with tool use, and multimodal document reasoning. SkillOpt was measured against baselines including no-skill, human-written skills, and one-shot LLM-generated skills, plus advanced prompt optimization and skill evolution methods like Trace2Skill, TextGrad, GEPA, and EvoSkill. SkillOpt dominated across all 52 evaluated combinations. With frontier models, it delivered an average absolute improvement of +23.5 points against the no-skill baseline on GPT-5.5. It also beat a hypothetical oracle baseline that cherry-picks the best competing method for every problem, which is a strong signal that SkillOpt’s approach is not just “sometimes better,” but consistently more effective across setups.

Zoom out and the second-order implications for decision-makers get sharper. First, SkillOpt produces compact, transferable skill artifacts that can adapt agents to new domains without rewriting everything from scratch. Second, it reduces the operational risk of automated skill updates because changes are gated by held-out validation and constrained by edit budgets. Third, it suggests a future where governance for agent behavior can be structured like training controls, not like open-ended prompt craftsmanship. For boards and leaders, the question becomes less “can we improve agent behavior” and more “can we do it fast enough without bleeding performance, reliability, and trust in production.” SkillOpt’s pitch is that you can upgrade skills automatically while keeping the model’s weights frozen, and in the enterprise world, that is exactly the kind of compromise that tends to get approved.