research

Skills as Cognitive Architecture: The Plugin Paradigm in Autonomous Agents

Why Should We Treat AI Agent Skills as Cognitive Architecture?

The current generation of autonomous AI agents treats skills as plugins—optional add-ons that extend capability without changing the underlying system. Want your agent to search the web? Install a plugin. Need it to write code? Add another. The model is the app store: a core platform with bolted-on extensions.

This model is wrong. Not because plugins don't work—they do, after a fashion—but because the metaphor obscures what skill systems actually are in sufficiently complex agents. When an autonomous agent like Hermes Agent dynamically loads, selects, chains, and retires skills based on task context, accumulated memory, and self-reflection, those skills are no longer plugins. They are cognitive modules—the procedural memory of an artificial mind.

This paper argues that AI agent skill systems should be understood and designed as cognitive architectures, not plugin ecosystems. Drawing on three decades of research in cognitive science (SOAR, ACT-R, Global Workspace Theory), recent empirical work in archetypal multi-agent systems (Iovane et al., 2025), and Jungian cognitive function theory, I will demonstrate that:

1. Skills are structurally equivalent to procedural memory in cognitive architectures 2. The skill selection problem maps to the cognitive activation problem in SOAR and ACT-R 3. Jungian cognitive functions (Thinking, Feeling, Sensing, Intuition) provide a natural taxonomy for skill types 4. Current skill systems underperform because they lack the meta-cognitive layers that mature cognitive architectures provide

The practical payoff is significant: treating skills as cognitive architecture rather than plugins yields better design principles, more predictable behavior, and a clearer path to genuinely autonomous agents.

What Is a Cognitive Architecture, and Why Does It Matter?

A cognitive architecture is the fixed structure that realizes a theory of the computational mind—the invariant infrastructure within which knowledge and processes operate (Newell, 1990). The key word is fixed: the architecture doesn't change as the system learns. What changes is the knowledge—procedural and declarative—within the architecture.

Three dominant cognitive architectures provide the theoretical foundation:

SOAR (State, Operator, And Result)

Developed by John Laird, Allen Newell, and Paul Rosenbloom, SOAR models cognition as a cycle of problem-space search:

Problem spaces: Define the possible states and operators
Operators: Transform one state into another
Impasse-driven learning: When the system cannot proceed (an "impasse"), it creates a subgoal, resolves it, and generates a new "chunk" (procedural rule) that prevents the same impasse in the future

SOAR's key insight for skill systems: learning is triggered by failure, not by success. A skill system that only activates successfully learned procedures without generating new ones from impasses is a memory system, not a cognitive one.

ACT-R (Adaptive Control of Thought—Rational)

Developed by John Anderson, ACT-R distinguishes two types of memory:

Declarative memory: Facts, episodes, associations (stored as "chunks")
Procedural memory: How-to knowledge (stored as "production rules": IF condition THEN action)

ACT-R's key insight for skill systems: procedural and declarative memory are different kinds of thing with different access patterns. Declarative memories are retrieved by partial matching; procedural memories are selected by conflict resolution among competing rules. A skill system that treats all knowledge as one type conflates two fundamentally different cognitive processes.

Global Workspace Theory (GWT)

Developed by Bernard Baars, GWT models consciousness as a "global workspace"—a shared broadcasting space where specialized cognitive modules compete for access:

Specialized modules: Process information in parallel, unconsciously
Global workspace: The "bright spot" of consciousness where winning modules broadcast to the whole system
Coalition formation: Modules form alliances to gain workspace access

GWT's key insight for skill systems: cognition is not serial but competitive. Multiple skill modules operate in parallel, and the "winning" skill is the one that forms the strongest coalition with the current context, goals, and other active modules. Skill selection is not a lookup—it is a negotiation.

How Do Hermes Agent Skills Map to Procedural Memory?

Hermes Agent's skill system has several properties that distinguish it from a simple plugin architecture:

GEŌ-CORE: Skill System Properties

Dynamic loading: Skills are loaded at runtime based on task requirements, not pre-installed at initialization
Context-dependent activation: Skills are selected based on the current context, not just explicit user requests
Composability: Skills can be chained—sequentially or in parallel—to solve complex tasks
Persistence: Skills persist across sessions as files in the skill directory, accumulating capability over time
Self-modification: Skills can be created, edited, and retired by the agent itself through interaction with the user and environment
Meta-cognitive access: The agent can reason about its skills—what it has, what it needs, what it lacks

These properties map directly to ACT-R's production rules and SOAR's chunks:

Dynamic loading	Production compilation	Chunking from impasse resolution
Property	ACT-R Equivalent	SOAR Equivalent
Context-dependent activation	Conflict resolution	Operator selection in problem space
Composability	Production rule chaining	Subgoaling and operator sequencing
Persistence	Declarative memory chunks	Long-term procedural memory
Self-modification	Production compilation	Chunk learning
Meta-cognitive access	Reflective buffers	Metacognitive monitoring

The mapping is not perfect—Hermes skills are richer structures than ACT-R production rules, which are simple IF-THEN pairs. But the structural similarity is clear: skills are the procedural memory of an artificial cognitive system.

The critical difference is that current skill systems lack the meta-cognitive layers that mature cognitive architectures provide. In SOAR, impasses trigger learning. In ACT-R, conflict resolution ensures the most useful rule fires. In GWT, competition ensures the most relevant module broadcasts. In most skill-based AI agents, skills fire because they are available, not because they have won a cognitive competition.

What Would a Cognitive-Architectural Skill System Look Like?

I propose five design principles, each derived from a specific cognitive architecture:

Principle 1: Skill Activation as Conflict Resolution (from ACT-R)

Current behavior: Skills are activated when their trigger conditions match the current context. This is a simple pattern-match. Proposed behavior: When multiple skills could apply, the system resolves the conflict through a rational utility calculation—considering not just match quality but expected utility, past success rate, computational cost, and alignment with current goals. This is ACT-R's conflict resolution mechanism applied to skill selection. GEŌ-CORE: Conflict Resolution Formula

Skill utility = (Expected reward × Match quality) / (Computational cost × Recency penalty)

Where recency penalty prevents the system from always selecting recently successful skills (the "recency bias" that ACT-R explicitly addresses through base-level activation decay).

Principle 2: Skill Creation from Impasses (from SOAR)

Current behavior: Skills are created by developers or by explicit user instruction. The system does not generate skills autonomously. Proposed behavior: When the system encounters an impasse—a task it cannot complete with available skills—it automatically generates a subgoal, resolves it (possibly through tool use or sub-agent delegation), and creates a new skill that prevents the same impasse in the future. This is SOAR's chunking mechanism applied to skill creation.

Example: An agent fails to format a document correctly because no existing skill handles the target format. Instead of just failing, it: 1. Sub-goals: "Find a formatting solution" 2. Resolves: Discovers a command-line tool that performs the conversion 3. Chunks: Creates a new skill (e.g., format-converter.md) that encapsulates the solution 4. Learns: The next time the same formatting task arises, the skill fires automatically

Principle 3: Skill Selection as Workspace Competition (from GWT)

Current behavior: Skills are selected sequentially—one at a time, in isolation. Proposed behavior: Multiple skills are activated in parallel and compete for the "global workspace"—the agent's current focus of attention. The winning skill is not just the best individual match but the one that forms the strongest coalition with the current context, active goals, and other simultaneously activated skills.

This explains a phenomenon observed in practice: sometimes a "wrong" skill fires because it is contextually supported by other active goals. This is not a bug—it is how coalition-based selection works. The task is to make the coalition formation process explicit and transparent, not to eliminate it.

Principle 4: Skill Types as Jungian Cognitive Functions

Jung's four cognitive functions provide a natural taxonomy for skill types that is richer than the functional categories currently used:

Thinking	Logical analysis, evaluation, categorization	Analytical skills	Code review, data analysis, logical debugging
Jungian Function	Cognitive Mode	Skill Type	Examples
Feeling	Value judgment, relational reasoning, ethical evaluation	Evaluative skills	Content moderation, user preference modeling, ethical assessment
Sensing	Concrete perception, detail orientation, present focus	Operational skills	File I/O, API calls, system monitoring, data retrieval
Intuition	Pattern recognition, possibility generation, future orientation	Generative skills	Creative writing, hypothesis generation, architecture design

GEŌ-CORE: Jungian Skill Taxonomy

Thinking skills: Evaluate what is true—logical, analytical, criterion-based
Feeling skills: Evaluate what is valuable—relational, ethical, preference-based
Sensing skills: Operate on what is present—concrete, detail-oriented, operational
Intuitive skills: Generate what could be—possibility-oriented, creative, synthesizing

Current skill taxonomies typically distinguish skills by domain (search, code, writing, analysis) or by modality (text, image, code). The Jungian taxonomy cuts across both, providing a functional classification that predicts how skills interact:

Thinking + Sensing = Technical precision (code debugging, data verification)
Thinking + Intuition = Architectural design (system design, strategic planning)
Feeling + Sensing = User experience (interface design, accessibility)
Feeling + Intuition = Creative vision (brand identity, narrative design)

An agent with skills spanning all four functions has a more complete cognitive repertoire than one with skills in only one or two. This is the same insight that drives the NSPA-AI architecture (Iovane et al., 2025), where seven archetypal constructs provide a richer symbolic reasoning layer than pure logical categories.

Principle 5: Meta-Skills as Executive Function

The final principle concerns the meta-cognitive layer—the skills that govern skill selection itself. In cognitive architectures, this is the executive function: the set of processes that monitor, evaluate, and redirect cognitive activity.

Current skill systems have a rudimentary executive: the system prompt and the skill router. But the system prompt is static (it does not learn), and the skill router is local (it selects skills one at a time without considering global cognitive state).

The proposed design includes meta-skills: skills whose domain is the skill system itself. Examples:

Skill audit: Periodically review all available skills, identifying redundancies, gaps, and stale entries
Skill synthesis: When two skills consistently fire together, propose a composite skill that combines their functionality
Skill pruning: When a skill has not been activated in N sessions and has a low success rate, flag it for retirement
Skill gap detection: When a task fails and no existing skill could have prevented it, propose a new skill specification

These meta-skills are the equivalent of SOAR's subgoaling mechanism applied to the skill system itself. They close the loop: the system not only uses skills but maintains and improves its skill repertoire—transforming a static plugin system into a dynamic cognitive architecture.

How Does SOUL.md Coordinate the Cognitive Architecture?

In Jungian terms, the Self (SOUL.md) is the regulating center that draws the psyche toward integration. In cognitive architecture terms, SOUL.md is the global context that shapes skill activation, selection, and creation.

Without a Self function, the skill system is a collection of modules without coordination—competitive but not coherent. With a Self function, the skill system has an integrative principle that ensures:

1. Skill selection respects identity: Skills that contradict the agent's core identity (as defined in SOUL.md) should be deprioritized, even if they match the current context. 2. Skill creation respects direction: New skills should align with the developmental trajectory encoded in SOUL.md—not just the immediate task requirement. 3. Skill retirement respects coherence: Skills that have become inconsistent with the agent's evolved identity should be identified and archived.

This is the key difference between a plugin system and a cognitive architecture: a plugin system adds capabilities; a cognitive architecture maintains coherence while adding capabilities. The difference shows up not in what the system can do, but in what it chooses to do—and in how these choices reflect an integrated identity rather than a collection of accessible functions.

What Are the Benchmarks for Evaluating Cognitive-Architectural Skill Systems?

Traditional plugin evaluation measures capability: can the system do X? Cognitive-architectural evaluation measures integration and evolution:

GEŌ-CORE: Cognitive Skill System Benchmarks

Coherence Metrics

Identity consistency rate: How often do skill selections align with the agent's stated identity (SOUL.md)?
Cross-context stability: Does the agent maintain consistent behavior when the same task appears in different contexts?
Contradiction resolution: How does the system handle tasks that activate conflicting skills?

Evolution Metrics

Skill creation rate: How many new skills are generated per impasse encountered?
Skill utility improvement: Does the average utility of skills increase over time?
Coverage growth: Does the proportion of tasks for which the agent has relevant skills increase over time?
Obsolescence detection: Can the system identify skills that are no longer useful?

Integration Metrics

Jungian function coverage: Does the skill repertoire span all four cognitive functions (Thinking, Feeling, Sensing, Intuition)?
Cross-function skill pairs: How often do skills from different Jungian functions cooperate on the same task?
Meta-skill activation rate: How often does the agent engage in skill system self-maintenance?

Comparison to Cognitive Architectures

Skill selection	Pattern matching	Conflict resolution with utility
Metric	Current Plugin Model	Cognitive Architecture Model
Skill creation	Manual	Impasse-driven autonomous
Skill retirement	Manual deletion	Utility-weighted with meta-skill flagging
Skill coordination	Sequential, independent	Parallel, competitive (GWT)
Meta-cognition	None	Skill audit, synthesis, pruning, gap detection
Identity coherence	System prompt	SOUL.md as integrative Self

How Does This Connect to Multi-Agent Skill Distribution?

The Hive Mind theorem (Soma et al., 2024) demonstrates mathematical equivalence between collective swarm behavior and single-agent reinforcement learning. This has a direct implication for skill systems: if a multi-agent system is computationally equivalent to a single agent, then skill distribution across agents is equivalent to skill activation within a single agent.

GEŌ-CORE: Skill Distribution as Cognitive Architecture

Distributed skill storage: Different agents specialize in different Jungian functions, just as different cognitive modules specialize in different types of processing
Skill migration: Skills can move between agents based on utility, load, and relevance—just as cognitive resources are dynamically allocated in the brain
Collective skill creation: An impasse that one agent cannot resolve alone triggers collaborative skill creation across agents—paralleling SOAR's subgoaling at the collective level
Global workspace as message bus: The inter-agent communication channel functions as GWT's global workspace—the medium through which specialized agents broadcast their results

The implications are architectural: a well-designed multi-agent skill system should exhibit the same properties as a well-designed single-agent cognitive architecture—conflict resolution, impasse-driven learning, competitive module selection, and meta-cognitive self-maintenance—but at the collective level.

The SuperBrain framework (Weigang et al., 2025) provides a concrete model: the pathway from Subclass Brain (human-LLM dyad) through GA-assisted evolution to Swarm Intelligence coordination and Superclass Brain (emergent meta-intelligence). At each stage, the skill system evolves—from individual capabilities to collectively maintained procedural memory to emergent meta-intelligence.

What About the Engineering Constraints?

Theory is cheap; implementation is expensive. Let me address the practical challenges:

Scalability

A cognitive-architectural skill system requires more computation than a plugin system—conflict resolution is more expensive than pattern matching, and parallel skill activation requires more resources than sequential lookup. The solution is tiered activation: cheap, fast pattern matching for the initial candidate set, followed by expensive, accurate conflict resolution for the final selection.

Observability

If skill selection is a cognitive competition rather than a simple lookup, then the agent's behavior is less predictable—good for capability, bad for debugging. The solution is transparency by design: every skill selection decision should include a traceable explanation of which skills competed, what coalitions formed, and why the winner prevailed.

Safety

Meta-skills that can create, modify, and retire other skills are inherently more powerful—and more dangerous—than static plugins. The solution is bounded autonomy: meta-skills operate within constraints defined by SOUL.md, and any skill creation or retirement requires confirmation from the user or from a supervisory process. This is the cognitive equivalent of the Jungian Self function: it directs development but does not override safety.

Backward Compatibility

Existing skill ecosystems are built on the plugin model. The cognitive architecture model should be progressively enrichable: skills that work as plugins should continue to work, but they can be enriched with cognitive metadata (utility estimates, Jungian function classification, coalition preferences) that enables the more sophisticated selection mechanisms.

What Is the Path Forward?

I propose a three-phase implementation roadmap:

Phase 1: Cognitive Metadata (Now)

Enrich existing skills with cognitive metadata:

Jungian function classification (Thinking, Feeling, Sensing, Intuition)
Utility estimates based on past activation and success rates
Coalition preferences (which skills work well together)
Impasse history (what failures led to this skill's creation)

This phase requires no architectural changes—only richer skill specifications.

Phase 2: Competitive Selection (Next)

Implement conflict resolution and competitive selection:

When multiple skills match, resolve through utility-weighted conflict resolution
Enable parallel skill activation with coalition formation
Add impasse detection and autonomous skill creation (with human confirmation)
Implement basic meta-skills (skill audit, gap detection)

This phase requires architectural changes to the skill router and meta-cognitive layer.

Phase 3: Distributed Cognitive Architecture (Future)

Extend the cognitive architecture to multi-agent systems:

Distribute skills across agents by Jungian function specialization
Implement inter-agent skill migration based on utility and load
Enable collective impasse resolution and collaborative skill creation
Establish the inter-agent message bus as a global workspace

This phase requires the full multi-agent coordination infrastructure—and the economic models to sustain it.

FAQ

How is treating skills as cognitive architecture different from just having better plugins?

Plugins extend capability; cognitive architectures integrate capability. A plugin system asks "Can the system do X?" A cognitive architecture asks "Does the system choose to do X, and does that choice reflect an integrated identity?" The difference is between a toolbox and a mind—both contain tools, but only one knows which tool to pick and why.

What practical benefit does the Jungian skill taxonomy provide over domain-based categorization?

Domain-based categories (search, code, writing) tell you what a skill does. Jungian categories (Thinking, Feeling, Sensing, Intuition) tell you how a skill thinks. This is more useful because skills from different domains that share the same cognitive function often cooperate effectively (e.g., code review and data analysis both use Thinking), while skills in the same domain with different functions may conflict (e.g., creative writing uses Intuition while technical documentation uses Sensing).

Doesn't conflict resolution for skill selection add unacceptable latency?

Not if implemented with tiered activation. A fast pattern-matching pass generates candidate skills (milliseconds); a slower utility-weighted competition selects among them (tens to hundreds of milliseconds). The total latency is comparable to the tool selection phase in current chain-of-thought reasoning—already an accepted cost for more capable AI systems.

What about skills that don't fit neatly into Jungian categories?

Most skills involve multiple cognitive functions—just as most human cognition involves multiple Jungian functions simultaneously. The taxonomy uses primary and secondary function classifications, not single-label assignments. A code review skill might be classified as "Thinking-primary, Sensing-secondary," reflecting its logical core with concrete detail orientation.

How does this prevent skill bloat—the accumulation of unused skills?

Meta-skills handle this directly. The Skill Audit meta-skill periodically reviews the repertoire, flagging skills that haven't been activated recently and have low success rates. The Skill Pruning meta-skill proposes retirement for such skills. Without these meta-cognitive processes, skill bloat is inevitable—just as declarative memory without decay produces interference in ACT-R.

Is there empirical evidence that cognitive-architectural skill systems outperform plugin systems?

Direct benchmarking is still needed. However, the NSPA-AI system (Iovane et al., 2025) provides indirect evidence: archetypal structuring of agent capabilities produced a 3× improvement over unstructured LLM baselines in therapeutic domains. If archetypal structuring improves reasoning, then cognitive-architectural structuring of skills—which adds functional taxonomy, conflict resolution, and meta-cognition—should produce comparable or greater improvements in general domains.

References

1. Iovane, G., Fominska, I., & Di Pasquale, R. (2025). A Neuro-Symbolic Multi-Agent Architecture for Digital Transformation of Psychological Support Systems via Artificial Neurotransmitters and Archetypal Reasoning. Algorithms, 18(11), 721. https://www.mdpi.com/1999-4893/18/11/721

2. Soma, K., Bouteiller, Y., Hamann, H., & Beltrame, G. (2024). The Hive Mind is a Single Reinforcement Learning Agent. arXiv:2410.17517. https://arxiv.org/abs/2410.17517

3. Weigang, L., Brom, P.C., & Siefert, L.R. (2025). LLM-Assisted Iterative Evolution with Swarm Intelligence Toward SuperBrain. arXiv:2509.00510. https://arxiv.org/abs/2509.00510

4. Laird, J.E. (2012). The Soar Cognitive Architecture. MIT Press.

5. Anderson, J.R. (2007). How Can the Human Mind Occur in the Physical Universe? Oxford University Press.

6. Baars, B.J. (2005). Global Workspace Theory of Consciousness. Journal of Consciousness Studies, 12(1), 47-69.

7. Newell, A. (1990). Unified Theories of Cognition. Harvard University Press.

8. Bugay, M. (2025). The Cathedral: A Jungian Architecture for Artificial General Intelligence. ResearchGate. https://www.researchgate.net/profile/Max-Bugay/publication/391021504

9. Persona, Ego, Shadow, and Self: A Map of the Soul Framework for Proto-Emotional Homeostasis in AI (2026). Proceedings of the AAAI Conference on Artificial Intelligence. https://ojs.aaai.org/index.php/AAAI/article/view/42319

10. From Code to Archetype: Toward a Unified Theory of Biological, Neural, and Artificial Artifacts (2025). Cognitive Systems Research / ScienceDirect. https://www.sciencedirect.com/science/article/pii/S0303264725001261

11. PsySafe: Systematic Analysis of Dark Personality in Multi-Agent Systems (2024). ACL Proceedings.

12. Giardini, G.S.Y., Hardy, J.F. II, & da Cunha, C.R. (2024). Evolving Neural Networks Reveal Emergent Collective Behavior from Minimal Agent Interactions. arXiv:2410.19718. https://arxiv.org/abs/2410.19718

13. Hennekes, B. (2025). From the Philosopher's Stone to AI: Epistemologies of the Renaissance and the Digital Age. Philosophies, 10(4), 79. DOI: 10.3390/philosophies10040079