Multi-Skill Automated Optimization

• Multi-Skill Automated Optimization is an iterative paradigm that refines explicit, reusable skill artifacts like structured documents and folders.
• It leverages methods such as textual gradient descent, evolutionary search, and reinforcement learning to improve quality and auditability without modifying backbone parameters.
• MS-AO supports multi-skill coordination by optimizing individual skills, bundles, or meta-skills, enabling effective retrieval and composition across diverse tasks.

Searching arXiv for papers directly relevant to multi-skill automated optimization and the named frameworks. Multi-Skill Automated Optimization (MS-AO) denotes a class of methods that automatically improve reusable agent skills, skill bundles, or skill libraries through iterative evaluation and revision rather than manual prompt engineering or weight updates. Across recent work, the optimized objects are typically externalized artifacts—such as structured skill documents, skill folders, memory scaffolds, orchestration specifications, or skill bundles—rather than latent policies. The defining pattern is closed-loop optimization over multiple reusable capabilities: systems induce or edit skills from experience, evaluate them with execution- or judge-based signals, retain or refine the strongest variants, and deploy them through retrieval, composition, or portfolio selection on new tasks. Recent arXiv work instantiates this pattern in several forms, including per-task skill induction with library retrieval (Li et al., 9 May 2026), modular memory-skill optimization (Wu et al., 1 Jul 2026), multi-objective bundle evolution (Gong et al., 10 Apr 2026), execution-free optimization-skill libraries for code efficiency (Wang et al., 29 Mar 2026), population-based reasoning-skill evolution (Hwang et al., 9 May 2026), portfolio-scale skill self-evolution (Tian et al., 30 Apr 2026), higher-order orchestration-skill evolution for multi-agent systems (Lin et al., 17 Jun 2026), skill-library discovery via failure analysis (Alzubi et al., 3 Mar 2026), and multi-objective Pareto optimization of skill artifacts under platform constraints (Tanjim et al., 19 May 2026).

1. Conceptual scope and defining characteristics

MS-AO is best understood as an optimization paradigm over modular capability artifacts. In the recent literature, a “skill” is not a single canonical object. In MIND-Skill, it is a “structured skill document” or SKILL.md containing sections such as Overview, When to Apply, Procedure, Key Patterns, and Common Pitfalls (Li et al., 9 May 2026). In EvoSkill, it is a filesystem-based skill folder comprising instructions, trigger metadata, and optional helper scripts (Alzubi et al., 3 Mar 2026). In Skills-Coach, a skill may include Skill.md, Readme.md, commands, parameters, and code/resources (Tian et al., 30 Apr 2026). In AutoMem, the optimized skill is memory management itself, exposed through file-system operations and a scaffold of prompts, schemas, and memory actions (Wu et al., 1 Jul 2026). In Skill-MAS, the optimized artifact is a structured Meta-Skill for task decomposition, agent engineering, and workflow orchestration (Lin et al., 17 Jun 2026). In MOCHA, the target is a SKILL.md artifact with a description field and an instruction body subject to routing and context constraints (Tanjim et al., 19 May 2026).

This suggests a broad but coherent definition. MS-AO concerns systems that optimize reusable procedural artifacts governing agent behavior, especially when those artifacts are modular, inspectable, and deployable across tasks. A plausible implication is that MS-AO occupies a middle layer between model-level adaptation and one-shot prompting: it changes the agent’s operational knowledge without modifying backbone parameters.

A second defining characteristic is that optimization is usually externalized and explicit. The papers do not primarily optimize hidden neural policies. Instead, they optimize prompts, skill documents, memory scaffolds, skill bundles, or orchestration specifications through search, reflection, textual gradients, evolutionary selection, or reinforcement learning over external skill generators (Li et al., 9 May 2026, Tanjim et al., 19 May 2026, Vishe et al., 10 May 2026). This externalization is repeatedly justified by portability, auditability, black-box compatibility, and the ability to work with closed-source frontier models (Vishe et al., 10 May 2026, Yim et al., 17 Mar 2026).

A third characteristic is multi-skill structure. The “multi” in MS-AO does not always mean a single optimizer jointly controls a tightly coupled set of skills. In some systems, optimization is independent per skill or per task, but the resulting library is used jointly through retrieval or composition. MIND-Skill learns one optimized skill per training task and retrieves top-$K$ skills at inference, with default $K=3$ (Li et al., 9 May 2026). EffiSkill retrieves three candidate sets of Operator Skills and composes them into plans via Meta Skills (Wang et al., 29 Mar 2026). SkillMOO directly optimizes a bundle of multiple coding skills as a single search object (Gong et al., 10 Apr 2026). Skill-MAS optimizes a Meta-Skill whose internal modules encode multiple orchestration sub-skills (Lin et al., 17 Jun 2026). Thus, “multi-skill” ranges from portfolio-scale optimization to explicit bundle optimization to higher-order orchestration over multiple sub-skills.

2. Skill representations and optimization targets

The recent literature converges on structured, reusable, natural-language-centered representations. MIND-Skill adopts the Anthropic-style view of skills as structured, human-readable artifacts and concretely produces a SKILL.md file containing procedural conventions and workflow patterns rather than raw trajectories or executable programs (Li et al., 9 May 2026). EffiSkill makes this structure explicit with two card schemas. An Operator Skill Card contains metadata such as skill_id, type, language, name, description, family, tags, and triggers, plus body fields including When to use, Transformation steps, Expected complexity effect, Common pitfalls, When not to use, and Minimal example. A Meta Skill Card contains Overview, Core loop, Routing heuristics, Budgeting / control logic, and Decision checklist (Wang et al., 29 Mar 2026). This separation between mechanism skills and orchestration skills is a recurring theme in MS-AO.

Some frameworks optimize richer artifacts than text alone. Skills-Coach optimizes both instructions and code through separate instruction and code pathways (Tian et al., 30 Apr 2026). AutoMem optimizes both the “structure that supports the skill”—prompts, file schemas, validation logic, memory organization, and action vocabulary—and the “proficiency” of the model exercising that structure via filtered supervised finetuning of a dedicated memory specialist (Wu et al., 1 Jul 2026). ASDA generates skill files containing reasoning procedures, code templates, and worked examples, organized by financial subfield and error type (Yim et al., 17 Mar 2026). EvoSkill materializes proposals into structured, reusable skill folders with SKILL.md, trigger metadata, and helper scripts (Alzubi et al., 3 Mar 2026).

In some papers the optimized object is not a domain skill but a higher-order capability. AgentPSO treats each agent’s state as a natural-language reasoning skill instruction and updates a population of such skills using PSO-inspired memory and self-reflective directions (Hwang et al., 9 May 2026). Skill-MAS optimizes a Meta-Skill that governs automatic multi-agent system generation through modules for Task Decomposition, Agent Engineering, and Workflow Orchestration (Lin et al., 17 Jun 2026). AutoMem treats metamemory—knowing what to encode, when to retrieve, and how to organize knowledge—as an independently learnable cognitive skill (Wu et al., 1 Jul 2026). These cases broaden MS-AO from domain-specific procedural playbooks to reusable control skills.

The representation choices imply a practical criterion for MS-AO: a capability becomes optimizable once it is made explicit enough to log, critique, retrieve, and revise. This suggests why many papers emphasize files, Markdown, schemas, and action vocabularies rather than latent prompting alone.

3. Optimization loops and learning mechanisms

The central methodological distinction within MS-AO is how skill artifacts are improved. Several papers instantiate different but related optimization loops.

MIND-Skill uses a two-agent induction–deduction loop. For each task specification $t$ and successful trajectory $\tau$, an induction agent $\mathcal{A}_I$ with optimizable prompt $\mathcal{P}_I$ generates a skill $s$, and a frozen deduction agent $\mathcal{A}_D$ reconstructs a trajectory $\hat{\tau}$ using only $t$ and $K=3$0. Three losses are then computed: outcome loss, reconstruction loss, and rubric loss. Crucially, optimization is lexicographic rather than additive:

$K=3$1

with

$K=3$2

TextGrad then updates the induction prompt from textual critiques rather than directly editing the skill (Li et al., 9 May 2026). This is a skill-optimization loop centered on faithfulness, executability, and abstraction control.

SkillMOO treats a multi-skill bundle $K=3$3 as the search variable in a bi-objective problem:

$K=3$4

A solver agent evaluates a candidate bundle on coding tasks, an optimizer agent proposes edits based on failure traces, and archived bundles are ranked via NSGA-II survivor selection using non-dominated sorting and crowding-distance tie-breaking (Gong et al., 10 Apr 2026). This is the clearest direct instance of bundle-level MS-AO.

MOCHA formalizes skill optimization as Pareto search over multiple objectives. For candidate skill $K=3$5 with objective vector

$K=3$6

the Pareto-optimal set is

$K=3$7

Parent selection uses randomized Chebyshev scalarization

$K=3$8

with $K=3$9 and ideal point $t$0. Acceptance interpolates between hypervolume-contribution exploration and scalarization-consistent exploitation via an annealed threshold (Tanjim et al., 19 May 2026). In effect, MOCHA contributes a formal multi-objective search layer for skill artifacts.

AgentPSO optimizes a population of reasoning skills using a prompt-mediated PSO analogy. Agents have current skill $t$1, personal best $t$2, global best $t$3, and semantic velocity $t$4. A self-reflective direction

$t$5

is extracted from peer reasoning trajectories, then combined through prompt-based operators:

$t$6

$t$7

Validation accuracy $t$8 updates personal-best and global-best memories (Hwang et al., 9 May 2026). This is notable because it optimizes multiple individualized skills simultaneously without weight updates.

Skill-R1 contributes a reinforcement-learning view of recurrent skill evolution. A trainable skill generator $t$9 produces skills $\tau$0 conditioned on task context and prior history, while a frozen task model induces rollout groups scored by verifiable rewards. The global objective is

$\tau$1

Its key innovation is a bi-level advantage:

$\tau$2

$\tau$3

$\tau$4

which is used בתוך a clipped GRPO-style objective with KL regularization (Vishe et al., 10 May 2026). This directly addresses recurrent credit assignment in skill evolution.

Skill-MAS similarly optimizes a reusable Meta-Skill over rounds of multi-trajectory rollout and selective reflection. For task $\tau$5 and rollout $\tau$6, trajectories are

$\tau$7

Per-task uncertainty and difficulty are defined as

$\tau$8

Priority-driven task selection uses

$\tau$9

followed by elbow-based truncation (Lin et al., 17 Jun 2026). This supplies a targeted reflection budget over the most informative failures.

Taken together, these papers show that MS-AO does not prescribe a single optimizer. Instead, it encompasses textual gradient descent, evolutionary search, Pareto selection, prompt-based swarm optimization, RL over skill generators, and teacher-guided iterative refinement, so long as the loop explicitly improves reusable skill artifacts.

4. Evaluation signals, verification, and quality control

A recurrent concern in MS-AO is how to judge skill quality independently of a strong or weak executor. The papers offer several solutions.

MIND-Skill is the most explicit about disentangling dimensions of quality. Outcome loss measures whether reconstructed trajectories solve the task in the environment; reconstruction loss measures tactic-level discrepancy between source and reconstructed trajectories; rubric loss scores the skill document itself along five dimensions: ground-truth independence, actionability, transferability, completeness, and conciseness (Li et al., 9 May 2026). The paper emphasizes that existing methods lack such quality guarantees, and that rubric plus reconstruction prevent memorized or poorly documented skills from scoring well.

AutoMem reframes delayed memory errors as trajectory-review problems. Its scaffold optimization loop uses complete episode traces and memory directories to diagnose failure modes such as duplicate map entries, empty searches, or poor file organization. Acceptance is gated by actual average progression on fixed seeds, not by purely textual judgments (Wu et al., 1 Jul 2026). This suggests a broader MS-AO principle: when skill effects are delayed, complete trajectory review can be more informative than scalar terminal reward alone.

Skills-Coach introduces a Traceable Evaluation Module that scores original and optimized skills using pass rate, average score, standard task score, advanced task score, and error rate, with criterion-level reports and retain/discard decisions (Tian et al., 30 Apr 2026). Evaluation can be LLM-based or heuristic fallback, but the system emphasizes traceability through criterion-by-criterion evidence.

ASDA uses a dual-phase verification gate. Candidate expansions or repairs to skill files are accepted only if student performance on attributed coverage-gap or regression cases passes thresholds $\mathcal{A}_I$0 or $\mathcal{A}_I$1, with multiple regeneration attempts allowed (Yim et al., 17 Mar 2026). This amounts to modular regression testing over the skill library.

MOCHA’s evaluation is explicitly multi-objective. Correctness is paired with description compliance and body compliance, where compliance is scored as

$\mathcal{A}_I$2

for description limit $\mathcal{A}_I$3 and body limit $\mathcal{A}_I$4 (Tanjim et al., 19 May 2026). This turns platform constraints into optimized objectives rather than post hoc filters. A plausible implication is that library-scale MS-AO should often evaluate not just task quality but also deployability constraints such as context occupancy, routing length, and latency.

The importance of diagnosis quality is also highlighted by production evidence on description optimization. The paper on production skill description optimization shows that description rewriting can match manual tuning for resolving skill collisions, but cannot fix genuinely overlapping intended scopes. It identifies a large train-validation F1 gap as a diagnostic for cases that require architectural rather than textual intervention (Zhou et al., 29 Jun 2026). This is a cautionary result: optimization loops need not only acceptance metrics but also diagnostics for non-textual failure modes.

5. Multi-skill coordination, retrieval, and composition

MS-AO methods differ substantially in how multiple skills are used at inference.

MIND-Skill produces “a library with one skill per training task” and retrieves top-$\mathcal{A}_I$5 relevant skills, default $\mathcal{A}_I$6, based on the test task. The selected skills are jointly injected into the agent prompt (Li et al., 9 May 2026). The paper’s analysis of $\mathcal{A}_I$7 reports that $\mathcal{A}_I$8 underperforms, $\mathcal{A}_I$9 substantially improves performance, and $\mathcal{P}_I$0 can further improve scenario-level consistency on AppWorld-Normal SGC. This supports a modular “progressive-disclosure principle” rather than monolithic playbooks.

EffiSkill’s online stage is explicitly multi-skill. After execution-free diagnosis, it retrieves $\mathcal{P}_I$1 candidate sets of Operator Skills, each corresponding to a plausible optimization route, then composes each bundle into $\mathcal{P}_I$2 to $\mathcal{P}_I$3 optimization plans via Meta Skills (Wang et al., 29 Mar 2026). Retrieval is therefore bundle-based from the start, and plan composition is a distinct stage between retrieval and generation.

ASDA performs LLM-based routing over SKILL.md, selecting one or more skill files for a query. The example of a fixed-income question using both fixed_income/wrong_method_selection.md and common/missed_constraints.md shows limited joint composition (Yim et al., 17 Mar 2026). However, the paper also identifies over-reasoning caused by loading all skill files for a subfield as a bundle, indicating the difficulty of coarse routing.

Skill-MOO optimizes a bundle directly. Tasks begin with original bundles containing multiple skills—8, 6, and 5 skills respectively across the three SkillsBench tasks—and the optimizer edits the whole bundle through pruning, substitution, reordering, rewriting, and sometimes expansion (Gong et al., 10 Apr 2026). Here composition is not learned online; it is the object of offline search.

Skill-MAS operates at a higher level: its Meta-Skill composes decomposition rules, agent-role design, and workflow-topology strategies into automatically generated multi-agent systems (Lin et al., 17 Jun 2026). In this sense, it optimizes not a flat skill set but a reusable orchestrator over multiple interacting sub-agents and workflow patterns.

The literature is correspondingly careful about what remains unsolved. MIND-Skill explicitly notes the absence of global redundancy minimization, conflict resolution, or hierarchical composition across skills (Li et al., 9 May 2026). Skills-Coach likewise optimizes one skill at a time rather than jointly coordinating multiple skills (Tian et al., 30 Apr 2026). Skill-R1 evolves essentially one skill trajectory per instance rather than a persistent multi-skill library (Vishe et al., 10 May 2026). This suggests that MS-AO as currently realized often optimizes skills modularly first and leaves full joint coordination as an open problem.

6. Empirical evidence across domains

Recent papers provide strong but heterogeneous empirical support for MS-AO.

MIND-Skill reports strong held-out transfer on AppWorld and BFCL-v3. With Qwen3.5-122B-A10B as skill-generation model, it achieves AppWorld-Normal TGC 71.4 and SGC 55.4, AppWorld-Challenge TGC 51.8 and SGC 39.6, BFCL-v3 ACC 77.3, and average 59.1, outperforming ACE at 56.1, Trace2Skill at 55.1, and Skill-extract at 51.0 (Li et al., 9 May 2026). Its ablation shows removing reconstruction or rubric materially degrades performance, with reconstruction especially important on Challenge TGC.

AutoMem reports that optimizing memory alone—without modifying task-action behavior—improves base-agent performance roughly $\mathcal{P}_I$4–$\mathcal{P}_I$5 across Crafter, MiniHack, and NetHack. The full pipeline improves progression from 25.00 to 51.36 in Crafter, 7.50 to 30.00 in MiniHack, and 0.42 to 1.85 in NetHack relative to the memory-as-file-system baseline, bringing a 32B open-weight model close to frontier proprietary systems on these tasks (Wu et al., 1 Jul 2026). Scaffold optimization provides the largest gains, with memory-specialist training adding smaller but consistent improvements.

SkillMOO reports that, relative to the original static skill bundle, it improves pass rate by +131.2% on Task 1, +30.8% on Task 2, and +2.1% on Task 3 while reducing cost by -31.7%, -5.4%, and -19.4% respectively, and runtime by -23.6%, -8.0%, and -29.8% (Gong et al., 10 Apr 2026). Pattern analysis shows pruning and substitution dominate expansion as sources of improvement.

EffiSkill outperforms execution-free baselines on EffiBench-X across Python and C++ with both GPT-5-mini and Qwen3-Coder-30B-A3B-Instruct. It reports Python/C++ OPT@1 and OPT@8 improvements over the strongest baseline by 3.69 to 12.52 percentage points depending on setting (Wang et al., 29 Mar 2026). Ablations show that removing retrieval or multi-plan composition sharply degrades performance, especially for Qwen.

AgentPSO achieves average majority-vote accuracy 71.98 across DeepMath, MATH, AIME25, Minerva, and BBH, outperforming static prompting and debate-based baselines such as MAD, MoA, and DMAD on average (Hwang et al., 9 May 2026). Cross-benchmark and cross-model transfer results suggest evolved skills capture reusable reasoning procedures rather than benchmark-specific prompt hacks.

Skills-Coach reports large before/after gains on 48 skills in Skill-X. Average score improves from 0.378 to 0.84, pass rate from 33.59% to 88.02%, standard task score from 43.00% to 87.43%, and advanced task score from 32.71% to 81.61% (Tian et al., 30 Apr 2026). Gains are especially large for low-performing and code-inclusive skills, though the paper lacks strong external baselines.

Skill-MAS reports average improvements from 21.68 to 29.49 on Gemini-3.1-Flash, 19.64 to 27.55 on GPT-5.4-Nano, 32.61 to 38.41 on Qwen3.5-Plus, and 33.72 to 41.05 on DeepSeek-V4-Flash across four complex benchmarks, with moderate inference cost increases (Lin et al., 17 Jun 2026). Transfer across tasks and LLMs remains positive, though weaker than matched-source results.

EvoSkill shows exact-match improvement on OfficeQA from 60.6% to 67.9% and SealQA improvement from 26.6% to 38.7%, with zero-shot transfer of a SealQA-evolved skill to BrowseComp improving accuracy by 5.3% (Alzubi et al., 3 Mar 2026). The merged-library result on OfficeQA is especially relevant to multi-skill complementarity.

ASDA shows that structured multi-skill distillation can outperform monolithic text optimizers in financial reasoning. On FAMMA with Haiku 3.5 as student, arithmetic accuracy improves from 41.00 to 58.33 (+17.33), and non-arithmetic from 49.21 to 55.16 (+5.95), substantially exceeding GEPA and ACE (Yim et al., 17 Mar 2026).

MOCHA shows that explicitly multi-objective selection matters. Across six tasks, it achieves mean correctness 0.675 versus 0.628 for the strongest baseline, a 7.5% relative improvement, while discovering 3.6 Pareto-optimal variants on average versus 1.6 for baselines and improving hypervolume by 3.1% (Tanjim et al., 19 May 2026).

Finally, production evidence on skill-description optimization grounds the enterprise relevance of MS-AO. On a 9-skill production group-chat agent with 372 regression cases, automated description optimization achieves 79.2% average F1 versus 79.4% for manually tuned descriptions, within a 0.78% noise floor, while reducing per-skill engineering effort from 120 minutes to 3.8 minutes (Zhou et al., 29 Jun 2026). The strongest empirical lesson is that a single rewrite captures most of the benefit.

7. Limitations, caveats, and open problems

Several caveats recur across the literature and define the current frontier of MS-AO.

First, many methods depend on high-quality successful trajectories or labeled failures. MIND-Skill requires successful trajectories or equivalent references (Li et al., 9 May 2026). EvoSkill relies on supervised datasets and task-specific scorers (Alzubi et al., 3 Mar 2026). ASDA requires labeled domain QA data and a teacher capable of producing useful failure analyses (Yim et al., 17 Mar 2026). AutoMem requires long-horizon traces plus strong meta-LLMs for scaffold revision and training-data curation (Wu et al., 1 Jul 2026). This suggests that MS-AO is often data-efficient relative to finetuning but not data-free.

Second, most current systems optimize skills independently or semi-independently rather than solving full joint multi-skill coordination. Skills-Coach explicitly optimizes one target skill at a time (Tian et al., 30 Apr 2026). MIND-Skill learns one skill per task and relies on retrieval at deployment (Li et al., 9 May 2026). Skill-R1 is recurrent single-skill optimization rather than library-level skill management (Vishe et al., 10 May 2026). Even SkillMOO, though bundle-level, is evaluated on small bundles and short search budgets (Gong et al., 10 Apr 2026). A plausible implication is that current MS-AO methods are strongest at modular skill refinement and still immature at large-scale inter-skill arbitration.

Third, quality-control mechanisms remain largely heuristic or empirical rather than formal. MIND-Skill’s “robust quality guarantees” are procedural, not theorem-level guarantees (Li et al., 9 May 2026). Skill-MAS similarly provides empirical rather than formal monotonic improvement guarantees (Lin et al., 17 Jun 2026). AgentPSO’s PSO analogy is conceptually useful but mathematically loose in language space (Hwang et al., 9 May 2026). Skill-R1 provides a principled objective but still operates over binary verifiable rewards and limited generations (Vishe et al., 10 May 2026).

Fourth, routing and composition remain weak points. ASDA reports skill-induced over-reasoning when routing is too coarse (Yim et al., 17 Mar 2026). The production description-optimization paper shows that description rewriting cannot fix genuinely overlapping intended scopes (Zhou et al., 29 Jun 2026). MOCHA highlights that co-resident skills compete for context and routing budgets, suggesting that portfolio-level objectives matter even when optimizing one skill at a time (Tanjim et al., 19 May 2026).

Fifth, several methods may be sensitive to the quality of LLM judges or reviewers. MIND-Skill uses LLM judges for reconstruction and rubric scoring (Li et al., 9 May 2026). AutoMem’s selection of “good” memory decisions is meta-LLM mediated (Wu et al., 1 Jul 2026). Skills-Coach relies partly on LLM-based evaluation (Tian et al., 30 Apr 2026). These systems may therefore inherit judge bias or noisy critiques.

Sixth, cross-model and cross-domain portability is mixed. AgentPSO and Skill-MAS show encouraging cross-model transfer (Hwang et al., 9 May 2026, Lin et al., 17 Jun 2026). EffiSkill’s explicit skill cards support cross-backbone reuse in code optimization (Wang et al., 29 Mar 2026). But ASDA shows that skills distilled for one student can hurt a stronger student, indicating model-specific failure distributions (Yim et al., 17 Mar 2026). This suggests that portability depends strongly on how abstract versus executor-specific the skill artifacts are.

8. Relation to adjacent traditions

MS-AO sits at the intersection of several older and neighboring traditions but differs from each in important ways.

Compared with prompt optimization, MS-AO usually optimizes richer, modular artifacts. MOCHA is closest to prompt optimization but already emphasizes structured skill fields, platform constraints, and Pareto trade-offs (Tanjim et al., 19 May 2026). Skills-Coach, ASDA, and EvoSkill go further by optimizing modular skill packages or libraries rather than single prompts (Tian et al., 30 Apr 2026, Yim et al., 17 Mar 2026, Alzubi et al., 3 Mar 2026).

Compared with memory systems, MS-AO treats at least some forms of memory as skills rather than fixed modules. AutoMem is the clearest example: memory management becomes an explicit action space and trainable specialist, with scaffold/interface optimization separated from model proficiency (Wu et al., 1 Jul 2026). This suggests a broader reinterpretation of agent architecture components as separately optimizable skills.

Compared with agent architecture search, Skill-MAS and Logic-Skill Programming illustrate that orchestration and composition can themselves be optimized as skills. Skill-MAS does this non-parametrically with a Meta-Skill for MAS generation (Lin et al., 17 Jun 2026). Logic-Skill Programming, though outside the recent LLM-agent wave, provides a relevant older example of optimization-based sequencing of learned skills under logical and geometric constraints (Xue et al., 2024). Its objective

$\mathcal{P}_I$6

shows a different but compatible view of multi-skill optimization: planning over skill sequences and subgoals rather than text artifacts (Xue et al., 2024).

Compared with classical combinatorial multi-skill allocation, the older operations-research literature is relevant mainly as an analogy. Hybrid Ant Colony Optimization for MS-RCPSP formulates multi-skill worker assignment under time and cost constraints (Myszkowski et al., 2016), and task assignment in multi-skill spatial crowdsourcing formalizes coverage-based assignment under budgets and feasibility constraints (Cheng et al., 2015). These works optimize multi-skill resource allocation rather than LLM-agent skill artifacts, but they reinforce a broader interpretation: MS-AO can mean any automated optimization over multiple heterogeneous capabilities under coupled constraints.

9. Prospective directions

The papers collectively suggest several likely next directions for MS-AO research.

One direction is library-level optimization. Most current methods optimize per-skill or per-bundle artifacts locally, then retrieve them heuristically. Future systems will likely need explicit objectives over whole libraries, including redundancy, routing accuracy, conflict minimization, context occupancy, and portfolio diversity. MOCHA’s Pareto framing and hypervolume metrics provide one plausible foundation for this (Tanjim et al., 19 May 2026).

A second direction is stronger credit assignment across interacting skills. Skill-R1’s bi-level advantage is an important template, but it addresses only recurrent single-skill revision (Vishe et al., 10 May 2026). Extending this to skill routing, bundle composition, and persistent library updates remains open.

A third direction is scaffold–proficiency factorization beyond memory. AutoMem argues that many skills decompose into an interface/scaffold problem and a proficiency problem (Wu et al., 1 Jul 2026). This suggests a general MS-AO program: first optimize the representational substrate and action vocabulary of a skill, then train or refine the specialist that uses it.

A fourth direction is safer and more interpretable verification. MIND-Skill’s rubric dimensions, Skills-Coach’s traceable evaluation, and production train-validation gap diagnostics all point toward richer quality assurance for skill artifacts (Li et al., 9 May 2026, Tian et al., 30 Apr 2026, Zhou et al., 29 Jun 2026). This is likely to be especially important in regulated or enterprise domains.

A fifth direction is finer-grained and more reliable routing. ASDA’s over-reasoning failures and the enterprise description-optimization paper’s distinction between fixable collisions and genuine scope overlap both indicate that retrieval and routing are central bottlenecks (Yim et al., 17 Mar 2026, Zhou et al., 29 Jun 2026). Future MS-AO systems will likely need learned routers, traffic-aware evaluation, or adaptive context allocation policies.

Finally, a plausible implication is that MS-AO will increasingly become the practical adaptation layer for closed-source frontier LLMs. Many of the strongest papers—MIND-Skill, ASDA, Skill-R1, Skill-MAS, AutoMem—derive much of their appeal from improving behavior without touching backbone weights (Li et al., 9 May 2026, Yim et al., 17 Mar 2026, Vishe et al., 10 May 2026, Lin et al., 17 Jun 2026, Wu et al., 1 Jul 2026). If this trend continues, MS-AO may serve as the principal mechanism for domain specialization, memory shaping, orchestration refinement, and library maintenance in production agent systems.