Agents that can rewrite their own simulated worlds and distill those rewrites into reusable modules now eclipse raw model scaling as the dominant path to higher accuracy. Two independent systems released this month demonstrate a co‑evolutionary loop where workflows become skills and feedback loops sustain evolution without any extra training.
Previously, LLM agents constructed ad‑hoc workflows at inference time but discarded the resulting procedures once a task finished, while skill libraries were handcrafted or assembled offline [1]. This static pipeline prevented agents from capitalizing on successes they had already achieved, leaving long‑term capability growth to be driven solely by larger backbones.
Earlier attempts to evolve skills relied on a single‑turn question‑answer signal, which quickly ran out of trustworthy gradients and left multi‑turn defects invisible [2]. Without a persistent source of feedback the evolution process stalled after the first round of fixes.
FlowEvo reaches 85.6% accuracy on ALFWorld, a 26.4‑point lead over the strongest baseline, while consuming roughly one third of the tokens ["On ALFWorld, it reaches 85.6%, 26.4 points above the strongest baseline, while using roughly one third as many tokens." 2607.21596]. The system compiles each successful workflow into a callable skill, stores it in a persistent bank, and reuses it either directly or as context for new plans, turning transient traces into permanent capabilities.
Across ten base models ranging from 7 B to 671 B parameters, FlowEvo outperforms the ExpeL baseline in 49 of 50 model‑dataset comparisons, showing that the gain comes from skill accumulation rather than raw scaling [1]. The authors track downstream utility for each skill and suppress those that cause negative transfer, ensuring that the bank grows only with beneficial routines.
When compiled skills are made available for reuse, performance jumps from 38.8% to 80.6%, confirming that the feedback loop of skill compilation is the dominant factor ["Enabling skill reuse on top of compilation produces the dominant improvement (38.8% to 80.6%), confirming that the gain is driven by skill feedback rather than curation." 2607.21596]. This leap dwarfs gains from merely curating a larger static library, because each skill carries its execution trace and contextual cues.
SkillEvo lifts production‑skill success by 23 points over self‑reflection evolution and by 15.4 points over single‑turn QA‑driven evolution across six cloud‑service categories, nine skills, and ninety‑eight reference files ["Across six categories of cloud services, 9 production Skills, and 98 skill-reference files, SkillEvo surpasses self‑reflection‑based evolution by 23.0 points and single‑turn‑QA-driven evolution by 15.4 points." 2608.13120]. The framework replaces a scalar rejection gate with an active governance layer that repairs factual degradation and structural bloat, keeping the skill set healthy over many iterations.
Its multi‑turn simulation rewrites defects layer by layer—each follow‑up question both consumes feedback and creates fresh gradients—keeping the evolutionary signal alive without expanding model size ["Follow‑up questions expose defects layer by layer, so that every round of revision both consumes feedback and produces new feedback." 2608.13120]. By treating interaction as a continuous gradient generator, SkillEvo avoids the decay observed in single‑turn pipelines and sustains improvement across long dialogues.
The reported gains were achieved using a powerful GPT‑4o‑mini backbone; the paper does not evaluate whether smaller or open models would achieve similar token efficiency [1]. Moreover, maintaining a persistent skill bank introduces overhead for indexing and versioning, and mismanagement could lead to subtle interference between skills.
SkillEvo relies on simulated user feedback, where follow‑up questions are used to expose defects layer by layer; however, the paper does not claim that these simulations faithfully capture real‑world defects, which could limit performance in noisy production settings. If the feedback generator is biased or incomplete, the evolution gradient can drift toward spurious optimizations rather than true capability growth.
Teams can now prioritize building persistent skill banks and multi‑turn evaluation pipelines over chasing larger parameter counts, re‑running existing benchmarks with half‑sized models to achieve comparable scores. The co‑evolutionary recipe suggests that future agent performance will be defined more by the richness of their reusable skill ecosystems than by raw scaling.
If this loop holds, will the next generation of agents be judged by the size of their skill libraries rather than the size of their parameters?
