MUSE-Autoskill Agent Overview

• MUSE-Autoskill Agent is an autonomous, skill-centric architecture that integrates metacognitive learning, explicit skill lifecycle management, and hierarchical memory for continual adaptation.
• It dynamically creates, evaluates, and refines skills to boost task success rates, enhance robustness to distribution shifts, and enable effective cross-agent skill transfer.
• Empirical evaluations demonstrate gains of up to 15 percentage points in task performance, improved sample efficiency, and significant reductions in hallucination.

The MUSE-Autoskill Agent is an autonomous, skill-centric learning and planning architecture that integrates metacognitive mechanisms, explicit skill lifecycle management, and hierarchical, memory-driven continual adaptation. It enables LLM agents to autonomously acquire, refine, and compose skills in unknown environments—yielding sustained improvements in task-solving capability, robustness to distribution shift, and cross-agent skill transfer. This article provides a comprehensive technical synthesis of the MUSE-Autoskill paradigm, supported by recent experimental results and implementation blueprints (Valiente et al., 2024, Lin et al., 26 May 2026, Yang et al., 1 Mar 2026, Yang et al., 9 Oct 2025).

1. Architectural Foundations

MUSE-Autoskill unifies several core conceptual building blocks:

• Metacognitive Core: Adaptation is driven by explicit self-monitoring (competence awareness) and regulation (strategy selection/skill deployment), instantiated via learnable models that guide the agent’s planning and learning cycles.
• Skill Lifecycle Management: Skills are first-class assets external to the LLM, managed through iterative creation, memory accumulation, evaluation (tests and runtime outcomes), cataloguing/selection, and refinement pipelines. Skills are represented as structured artifacts (e.g., SKILL.md with versioning and tests) (Lin et al., 26 May 2026, Yang et al., 1 Mar 2026).
• Hierarchical Experience Memory: Multiple levels of memory—strategic (problem–solution), procedural (SOPs), tool use, and skill clusters—organize both atomic experiences and higher-level abstractions. These modules inject accumulated experience into decision contexts, biasing future plans without parameter updates (Yang et al., 9 Oct 2025).
• Meta-Controller and Autoskill Extensions: A meta-controller manages skill discovery (via competence-driven triggers and novelty detection), hierarchical macro-skill composition, and exploration–exploitation balancing using uncertainty-aware scheduling and continual memory integration (Valiente et al., 2024).

The system operates in a closed loop with the canonical ReAct-style plan–act–observe–reflect–memorize flow, extended to incorporate explicit competence estimation, modular skill banks, and dynamic skill evolution.

2. Metacognitive Skill Learning and Regulation

MUSE-Autoskill implements metacognition through two coupled modules:

• Self-Awareness Module: Continuously assesses the agent’s competence on current tasks or subtasks. In LLM-based variants, this is realized via a competence model $M_{sa}$ (e.g., transformer+MLP), mapping imagined or actual execution trajectories to scalar or probabilistic competence estimates.
• Self-Regulation Module: Selects or generates strategies (skills or composite policies) conditioned on these competence predictions. The regulator maximizes a mixed utility function:

$$U(\pi_i; s_t)\;=\;\mathbb{E}_{\tau\sim P(\cdot|\pi_i,s_t)}[C(\tau)]\;-\;\lambda\,\mathrm{Var}_{\tau}[C(\tau)]$$

where $\pi_i$ is a candidate skill, $C(\tau)$ is cumulative predicted competence over an imagined trajectory $\tau$ (Valiente et al., 2024). The tradeoff parameter $\lambda$ mediates risk sensitivity.

Strategy selection is implemented as a loop: candidates are simulated via chain-of-thought or world-model rollouts, scored by competence, and the highest-utility action/skill is selected for deployment. The system supports both pre-learned (atomic) and composed (macro) skills.

3. Skill Lifecycle: Creation, Memory, Evaluation, and Refinement

The agent maintains an external Skill Bank comprising versioned, modular skill artifacts. The full skill lifecycle includes:

• On-Demand Skill Creation: When no matching skill is found for a given subgoal, the skill_create tool is invoked, synthesizing a structured SKILL.md interface, implementation scripts, and tests via the LLM. Skills are registered only after all bundled unit tests pass. Failing tests initiate refinement via automatic patch generation and retry cycles (Lin et al., 26 May 2026).
• Skill-Level Memory and Update: Each skill appends timestamped logs, runtime outcomes, and test results to an append-only .memory.md file. This persistent experience is surfaced for future selection and adaptation.
• Structured Evaluation: Skills are evaluated both offline (unit test suites) and online (binary pass/fail per invocation), with test success rate (TSR) and runtime metrics used for continuous performance tracking:

$$\mathrm{TSR}(s) = \frac{1}{m} \sum_{i=1}^m \mathbf{1}[\text{run}(t_{s,i}) = \text{pass}]$$

• Refinement Loop: Skills failing tests or runtime checks are automatically patched, with up to a capped number of refinement cycles per attempt. Upon successful fixing, the skill version is incremented, and past memory logs may be reset to prevent legacy errors from affecting ongoing success rates (Lin et al., 26 May 2026).

The system supports periodic skill merge (for redundant interfaces) and pruning (low-utility skills) based on prior success rate and usage frequency metrics.

4. Hierarchical Memory and Autoskill Composition

MUSE-Autoskill employs a hierarchical memory structure:

• Strategic Memory ($\mathcal{M}_{strat}$): Stores compact, macro-level dilemma–strategy pairs, facilitating retrieval of high-level solution patterns (Yang et al., 9 Oct 2025).
• Procedural Memory ($\mathcal{M}_{proc}$): Structured as SOPs, these encode subtask-level, repeatable step sequences identified from accumulated trajectories.
• Tool Memory ($\mathcal{M}_{tool}$): Organizes both static tool descriptions and dynamic next-action hints, supporting efficient tool utilization and adaptation.
• Skill Memory ($$U(\pi_i; s_t)\;=\;\mathbb{E}_{\tau\sim P(\cdot|\pi_i,s_t)}[C(\tau)]\;-\;\lambda\,\mathrm{Var}_{\tau}[C(\tau)]$$0): (Autoskill extension) Abstracts and clusters multiple SOPs into higher-level skill prototypes, fostering compositionality and cross-domain skill induction. Clustering is performed periodically over embedding representations of observed SOPs (Yang et al., 9 Oct 2025).

Automated induction of new skills emerges via reflection on execution traces, unsupervised SOP clustering, and pattern extraction. These abstractions are registered as new skill prototypes—supporting both macro-level (hierarchical) and cross-task generalization.

5. Dynamic Skill Retrieval and Injection

At each decision point, the agent dynamically retrieves skills from the Skill Bank:

• Hybrid Retrieval: The selection mechanism combines dense embedding similarity and BM25 keyword retrieval to identify the most relevant skills to the current context or user query (Yang et al., 1 Mar 2026).
• Skill Context Injection: Retrieved skills are rendered as structured lists (with names, descriptors, interfaces, triggers, tags, examples) and provided in prompt context to the response generator, controlling which capabilities are activated.
• Bank Evolution and Drift Prevention: New skills undergo a merge/judge protocol against bank entries, with “add”, “merge”, or “discard” decisions handled by LLM-based agents. This ensures the skill bank remains durable and avoids uncontrolled drift.

This approach supports both individualized and cross-user skill personalization, as well as transparent transfer/export via standardized SKILL.md (or JSON/YAML) artifacts.

6. Empirical Evaluation and Comparative Performance

MUSE-Autoskill demonstrates performance gains across standard benchmarks:

• SkillsBench (51 real-world tasks): Incorporation of MUSE-generated skills yields a +15 percentage point lift in macro-averaged task success over the no-skills baseline for both Codex and Hermes agents, with MUSE achieving 68.40% vs. 53.19% (without skills) (Lin et al., 26 May 2026). On automatically generated skills (for 68.6% of tasks), MUSE achieves 87.94% accuracy, exceeding the human-authored skill ceiling.
• Cross-Agent Skill Transfer: Injecting MUSE-generated skills into Hermes increases performance from 47.89% to 58.40%, closing 79% of the gap to the human-skills reference, evidencing true portability of structured skill artifacts.
• Generalization and Continual Learning (TAC benchmark): On long-horizon productivity tasks, MUSE-Autoskill with hierarchical memory shows a ∼10–12% absolute gain in S_partial over baselines, and 41.14% final PCR—surpassing previous SOTA (Yang et al., 9 Oct 2025).
• Other Gains: Integrating per-skill lifecycle and dynamic injection leads to significant reductions in hallucination, improved style/personalization adherence, and measurable improvements in sample efficiency and success rates across domains (Yang et al., 1 Mar 2026).

7. Relation to Prior Architectures and Future Prospects

MUSE-Autoskill synthesizes cognitive-inspired metacognition (Valiente et al., 2024), memory-based continual learning (Yang et al., 9 Oct 2025), and explicit, compositional skill management (Lin et al., 26 May 2026, Yang et al., 1 Mar 2026), advancing beyond prior static or anthropic agent baselines:

Component	Static LLM Agents	Reflexion/ReAct	MUSE-Autoskill
Skill acquisition	N/A or prompt-only	Prompt+recall/reflection	On-demand, explicit
Skill memory	None (stateless)	Ad hoc history	Per-skill, append-only
Evaluation	Ad hoc runtime output	N/A	Unit tests + runtime
Lifecycle management	N/A	No	Full (create-manage-refine)
Skill transfer	N/A	N/A	Artifact-level sharing

A plausible implication is that further integration of metacognitive modules with unified skill lifecycle management will be essential to enable autonomous, robust, and continually improving agentic systems. Key directions include optimizing context management for large skill banks, compositional skill induction via unsupervised trajectories, and standardized interfaces for cross-agent interoperability (Valiente et al., 2024, Lin et al., 26 May 2026, Yang et al., 1 Mar 2026, Yang et al., 9 Oct 2025).