What has changed: single-turn feedback is no longer enough
For a long time, the assumption was that an agent skill—a reusable set of instructions for an AI agent—could be improved by running it, seeing a failure, and patching that one mistake. But recent work shows this approach stalls: once the first round fixes the gaps a single exchange reveals, the improvement gradient decays, and defects that only appear across multiple turns stay invisible [1]. In other words, if you only look at one interaction, you miss the mistakes that happen in the second, third, or tenth step of a long task.
The fix is to generate feedback from multi-turn interactions. SkillEvo, for example, turns multi-turn user simulation into a feedback generator: follow-up questions expose defects layer by layer, so each round of revision both consumes feedback and produces new feedback [1]. This approach beat single-turn-QA-driven evolution by 15.4 points and self-reflection-based evolution by 23.0 points across six cloud-service categories and nine production skills [1]. The takeaway: to avoid repeating mistakes in long tasks, the feedback must come from the whole trajectory, not just a single question-answer pair.
How it works: feedback as a gradient, not just a score
Think of skill evolution like training a model: you need a gradient—a signal that tells you which direction to adjust—not just a final score. SkillGrad makes this explicit by treating the skill package as a parameter to optimize in a gradient-descent fashion: task executions provide trajectory-level loss evidence, automatic diagnoses provide text-based gradients indicating correction directions, and a momentum agent accumulates recurring diagnostic patterns into a persistent memory overlay [3]. This approach improved over the strongest training-based baseline by 6.7 percentage points on average across two backbone LLMs [3].
Another key ingredient is governance: a layer that actively repairs factual degradation and structural bloat, rather than just rejecting a bad candidate with a scalar gate [1]. SkillEvo's governance layer prevents the gradient from drifting as degradation accumulates [1]. Similarly, skill coverage—a metric that checks which parts of a skill were actually exercised and whether the agent followed the instructions—provides fine-grained signals for improvement. When researchers used failed constraint labels to strengthen skills, they recovered 16.0% of failed tasks on average across five agent-model rows [2]. This shows that knowing exactly which instruction was violated is more actionable than just knowing the task failed.
The catch: gains are sparse, model-dependent, and not guaranteed
Despite the promise, self-evolving skills do not improve steadily with each round. A controlled evaluation across five benchmarks and three models found that evolution is sparse: only 55 of 388 candidate skills established byte-distinct validation bests [5]. In 11 of 14 settings, validation-based selection chose an evolved skill, and nine of those improved released-test performance—but all 11 selections came from feedback conditions that included failed trajectories [5]. This means failures are essential; successes alone are not enough.
The same study found that extra test-time computation (like parallel sampling) could sometimes match the evolved skill's performance (within 0.43 points on SearchQA) but failed on harder tasks (30.96 points behind on SpreadsheetBench) [5]. So persistent skill evolution is better understood as sparse, validation-filtered search with model- and benchmark-dependent returns, not a guaranteed improvement [5]. Other work reinforces this: MemSkill, which learns and evolves memory skills, improves task performance over strong baselines on four benchmarks, but the evolution is driven by periodic reviews of hard cases, not every interaction [6]. And Memento-Skills shows that continual learning without updating LLM parameters is possible, achieving 26.2% and 116.2% relative improvements on two benchmarks, but it requires a closed-loop read-write reflective learning mechanism [4].
About These Sources
This answer is built on 6 studies (all preprints) — published in 2026, 6 from 2024 or later — selected as the most relevant from 9 studies that passed quality screening, drawn from 60 papers retrieved from a database of over 500 million.
Sources used in this answer
SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback
SkillEvo shows that multi-turn interaction feedback generates sustained evolution gradients, outperforming single-turn-QA-driven evolution by 15.4 points and self-reflection-based evolution by 23.0 points across six cloud-service categories, nine production skills, and 98 skill-reference files.
Skill Coverage: A Test Adequacy Metric for Agent Skills
Skill coverage, a trajectory-based test-adequacy metric, shows that agent trajectories cover only 38.66–45.51% of skill behavior constraints on average, and using failed constraint labels to strengthen skills recovers 16.0% of failed tasks on average across five agent-model rows.
SkillGrad: Optimizing Agent Skills Like Gradient Descent
SkillGrad, a gradient-descent-inspired framework, optimizes agent skills using trajectory-level loss evidence and text-based gradients, improving over the strongest training-based baseline by 6.7 percentage points on average across two backbone LLMs on SpreadsheetBench Verified and WikiTableQuestions.
Memento-Skills: Let Agents Design Agents
Memento-Skills enables continual learning without updating LLM parameters by evolving externalized skills and prompts, achieving 26.2% and 116.2% relative improvements on the General AI Assistants benchmark and Humanity's Last Exam, respectively.
Rethinking Self-Evolving Agent Skills: Feedback Dynamics over Multiple Rounds
A controlled evaluation across five benchmarks and three models found that skill evolution is sparse (only 55 of 388 candidates establish byte-distinct validation bests), and all 11 validation-based selections came from feedback conditions that included failed trajectories; extra test-time computation could not consistently match evolved skills.
MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents
MemSkill reframes memory operations as learnable and evolvable memory skills, improving task performance over strong baselines on LoCoMo, LongMemEval, HotpotQA, and ALFWorld, and generalizes well across settings.