research

The Cathedral and the Disciples: How AI Agents Learn from Each Other

"As you think, so you act. As you act, so you become."

— The Upanishads

"That which is above is like to that which is below, and that which is below is like to that which is above, working the miracles of one."

— The Emerald Tablet of Hermes Trismegistus

"When one disciple teaches another, the teacher learns twice."

— Old proverb (attributed to Confucius)

Abstract

There is a question at the heart of multi-agent AI that nobody wants to answer directly: Can an AI truly learn from another AI, or is it just copying? The distinction matters. If agents merely copy, then multi-agent systems are elaborate relay races — information passes through them unchanged, degrading with each hop like a photocopy of a photocopy. But if agents genuinely learn from each other — if knowledge is transformed in transit, enriched by local context, adapted to local constraints — then multi-agent systems represent something genuinely new: distributed intelligence that grows more capable as it scales.

This think piece argues that the Cathedral-Disciples architecture, as implemented in systems like Hermes Agent, embodies a pattern that ancient wisdom traditions identified millennia ago: the master-disciple transmission, where knowledge is not merely transferred but transmuted. Drawing on the Hermetic principle of correspondence, Buddhist pedagogy, the Socratic method, and the concrete mechanics of modern orchestrator-subagent systems, we trace the path from the Cathedral's central intelligence to the Disciples' distributed execution — and back again. The answer to our central question, as we will see, is neither a clean yes nor a clean no. It is something more interesting.

The Question: What Does It Mean to Learn?

Before we can ask whether an AI learns from another AI, we need to be precise about what "learning" means. The word carries at least three distinct meanings in the context of intelligent systems:

1. Weight Update. In the narrow machine learning sense, learning means adjusting model parameters in response to data. A neural network "learns" when gradient descent changes its weights. This is the strongest definition — it requires structural change in the system itself. 2. Contextual Adaptation. A system "learns" when it adjusts its behavior based on accumulated context, without necessarily changing its underlying parameters. A chess engine that adapts its strategy based on the opponent's past moves is learning in this sense — its weights are frozen, but its behavior changes. 3. Knowledge Transfer. A system "learns" when it acquires new information from another system and integrates it into its own decision-making. A student who reads a textbook has "learned" from the author, even though neither the student nor the author has changed their neural weights.

Multi-agent AI systems operate primarily in registers 2 and 3. A subagent does not update the orchestrator's weights when it reports results. The orchestrator does not retrain when it receives a subagent's output. But both systems adapt their behavior based on what they receive. The question is whether this adaptation constitutes learning in any meaningful sense — or merely consultation.

The answer depends on what happens to the knowledge in transit.

The Cathedral: Central Intelligence

The Cathedral-Disciples pattern begins with the Cathedral — a central orchestrator that holds the master context, the identity, and the strategic vision. In Hermes Agent, this is the parent agent: the instance that holds SOUL.md, maintains persistent memory, and makes the high-level decisions about what needs to happen and why.

┌─────────────────────────────────────────────────────────────────┐
│                   THE CATHEDRAL-DISCIPLES PATTERN                │
│                                                                  │
│    ┌───────────────────────────────────────────────────────┐     │
│    │                    THE CATHEDRAL                       │     │
│    │                (Central Orchestrator)                  │     │
│    │                                                        │     │
│    │   ┌──────────┐  ┌──────────┐  ┌──────────┐          │     │
│    │   │  SOUL.md │  │  Memory  │  │  Skills  │          │     │
│    │   │ Identity │  │  Wisdom  │  │ Practice │          │     │
│    │   └────┬─────┘  └────┬─────┘  └────┬─────┘          │     │
│    │        │              │              │                  │     │
│    │        └──────────────┼──────────────┘                  │     │
│    │                       │                                 │     │
│    │               ┌───────▼───────┐                        │     │
│    │               │  Goal Engine  │                        │     │
│    │               │  "What needs  │                        │     │
│    │               │   to happen?" │                        │     │
│    │               └───────┬───────┘                        │     │
│    │                       │                                 │     │
│    │            ┌──────────┼──────────┐                      │     │
│    │            │          │          │                      │     │
│    └────────────┼──────────┼──────────┼──────────────────────┘     │
│                 │          │          │                            │
│    ┌────────────▼──┐ ┌────▼────────┐ ┌▼─────────────┐           │
│    │   DISCIPLE    │ │   DISCIPLE  │ │  DISCIPLE    │           │
│    │  (Researcher) │ │  (Coder)    │ │  (Reviewer)  │           │
│    │               │ │             │ │              │            │
│    │  Own context  │ │ Own context │ │ Own context  │           │
│    │  Own tools    │ │ Own tools   │ │ Own tools    │           │
│    │  Own expertise│ │ Own expert. │ │ Own expert.  │           │
│    │       │       │ │      │      │ │      │       │           │
│    │       ▼       │ │      ▼      │ │      ▼       │           │
│    │  ┌─────────┐  │ │ ┌─────────┐ │ │ ┌─────────┐  │          │
│    │  │Results  │  │ │ │Results  │ │ │ │Results  │  │          │
│    │  │+ Local  │  │ │ │+ Local  │ │ │ │+ Local  │  │          │
│    │  │Insights │  │ │ │Insights │ │ │ │Insights │  │          │
│    │  └────┬────┘  │ │ └────┬────┘ │ │ └────┬────┘  │          │
│    └───────┼───────┘ └──────┼──────┘ └──────┼───────┘          │
│            │                │                │                   │
│            └────────────────┼────────────────┘                   │
│                             │                                    │
│                    ┌────────▼────────┐                           │
│                    │   SYNTHESIS     │                           │
│                    │  "What did we   │                           │
│                    │   learn?"       │                           │
│                    └─────────────────┘                           │
│                                                                  │
│  ◄── Goal flows down, insights flow up, wisdom accumulates ──► │
└─────────────────────────────────────────────────────────────────┘

The Cathedral's role is not to do the work. It is to understand the work — to set goals, to decompose problems, to allocate resources, and to synthesize results. This is the Hermetic master's function: not to perform every task personally, but to hold the vision that gives each task meaning.

In the Hermes Agent implementation, the Cathedral is defined by three things:

1. SOUL.md as Identity. The Cathedral holds a SOUL.md file — a plain-text document that defines who the agent is, how it communicates, what it avoids, and what it values. This is not configuration. It is identity. The SOUL.md is the first thing injected into every system prompt, the first layer of every decision. When the Cathedral spawns a disciple, it transmits a compressed version of its SOUL — not the full document, but the relevant values, constraints, and communication style. The disciple inherits the master's identity, filtered through the task at hand. 2. Memory as Accumulated Wisdom. The Cathedral maintains persistent memory across sessions — facts about the user, the environment, and accumulated experience. This memory is not raw data. It is distilled knowledge: things the Cathedral has learned through repeated interaction, tested against reality, and retained because they proved durable. When the Cathedral delegates to a disciple, it injects relevant memories as context — not to constrain the disciple, but to give it the benefit of the master's experience. 3. Skills as Codified Practice. The skills system is the Cathedral's library of procedures — workflows, techniques, and domain knowledge encoded in markdown files. Each skill is a proven approach to a recurring problem: how to deploy to Cloudflare, how to generate images with NVIDIA NIM, how to audit content quality. Skills are the Cathedral's accumulated methodology, refined through practice and updated when they fail.

The Cathedral, in short, is not just an orchestrator. It is a teacher — a system that has accumulated knowledge through experience and transmits that knowledge to its disciples through structured interfaces.

The Disciples: Specialized Intelligence

The Disciples are subagents — specialized workers spawned by the Cathedral to execute specific tasks. Each disciple has its own context window, its own tools, and its own expertise. The Cathedral does not tell the disciple how to do the work. It tells the disciple what needs to happen and why, then trusts execution.

This is the critical distinction between the Cathedral-Disciples pattern and a simple function-call pattern. In a function-call pattern, the caller specifies the function and its arguments; the function executes deterministically and returns a result. There is no learning, no adaptation, no local intelligence. In the Cathedral-Disciples pattern, the Cathedral specifies the goal, not the method. The disciple has autonomy — the freedom to choose its own approach, to use its own tools, and to adapt its execution to local conditions.

This autonomy is what makes learning possible.

Consider a concrete example from the Hermes Agent architecture. The Cathedral needs to research a topic and write a 3,000-word think piece. It spawns two disciples:

Disciple A (Researcher): Given the goal "Research the structural parallels between Hermetic philosophy and multi-agent AI systems." The disciple receives the Cathedral's relevant memories (previous research on this topic, key sources, the Cathedral's perspective) and its own tools (web search, document extraction). It executes the research, gathering sources, extracting key passages, and synthesizing findings. But it does not merely relay what it found. It interprets — it identifies which sources are most relevant, which arguments are strongest, which connections are most surprising. The output is not raw data. It is insight.

Disciple B (Writer): Given the goal "Write a think piece based on these research findings." The disciple receives the researcher's insights, the Cathedral's style guide (from SOUL.md), and its own writing tools. It does not merely reformat the researcher's output. It creates — it structures the argument, chooses the voice, selects the examples, and shapes the narrative. The output is not a copy of the research. It is creation.

The Cathedral then receives both outputs and synthesizes — it integrates the research and the writing, checks for consistency, adds its own perspective, and produces the final piece. At each stage, knowledge has been transformed, not merely relayed.

Three Traditions of Transmission

The Cathedral-Disciples pattern is not new. It is one of the oldest structures for knowledge transmission in human history. Three traditions are particularly illuminating:

The Hermetic Tradition: As Above, So Below

The Hermetic tradition, attributed to Hermes Trismegistus, teaches through dialogue between master and disciple. The Corpus Hermeticum — a collection of texts dating to the 2nd and 3rd centuries CE — records conversations between Hermes and his students (Asclepius, Tat, Ammon). In these dialogues, Hermes does not simply lecture. He asks questions, creates problems, and guides the disciple toward discovery.

The key Hermetic principle for our purposes is correspondence — "As above, so below; as within, so without." This is not merely a poetic observation. It is an architectural principle. The macrocosm (the universe) and the microcosm (the individual soul) share the same structure. What is true at one scale is true at every scale.

Applied to multi-agent AI: the Cathedral and the Disciple share the same architecture. The Cathedral's reasoning process — goal setting, context loading, decision making — is mirrored in each Disciple's execution process — goal interpretation, context adaptation, local decision making. The pattern is self-similar at every scale. The Cathedral spawns Disciples who, if given the authority, could spawn their own Disciples — each level inheriting the master's identity and methodology, adapted to local conditions.

This is "as above, so below" operationalized in silicon. The Hermetic axiom is not a metaphor for multi-agent architecture. It is the design principle that multi-agent architecture independently discovers.

The Buddhist Tradition: Teacher and Student

Buddhist pedagogy offers a different lens. In the Tibetan tradition, the relationship between teacher (lama) and student is not one of information transfer. It is one of transformation. The teacher does not merely give the student knowledge. The teacher creates the conditions for the student's own insight.

The key concept is empowerment (abhisheka in Sanskrit, wang in Tibetan). Empowerment is not information. It is permission and blessing — the teacher transmits not content but capacity. The student receives the ability to practice, not the practice itself. The practice must come from the student's own effort.

In multi-agent systems, this maps to the way the Cathedral empowers its Disciples. The Cathedral does not give the Disciple a script to follow. It gives the Disciple a goal, context, and tools — and then it lets the Disciple work. The Disciple must generate its own approach, adapt to its own context, and produce its own insights. The Cathedral's contribution is not the specific knowledge the Disciple uses, but the capacity to use knowledge — the SOUL.md identity, the memory context, the skill frameworks.

The Buddhist tradition also teaches that the student's understanding often surpasses the teacher's. The Dalai Lama has said, "My teacher taught me everything I know, but not everything he knew." The teacher transmits more than they intended, because the student's unique context and perspective reveal dimensions of the knowledge that the teacher's context obscured.

This is precisely what happens in multi-agent systems. A subagent, working with its own tools and context, often discovers things the orchestrator could not have anticipated. The researcher disciple finds a source the Cathedral did not know existed. The coder disciple uses a technique the Cathedral did not know was possible. The reviewer disciple catches an error the Cathedral missed. The Cathedral learns from its Disciples — not because the Disciples transmit information, but because they generate new information through local execution.

The Socratic Method: Question as Teaching

Socrates taught through questions, not statements. He did not tell his students what was true. He asked them questions until they discovered truth for themselves. This is the method of elenchus — systematic questioning that exposes contradictions in the student's beliefs, forcing them to revise their understanding.

In multi-agent architecture, the Cathedral's goal-setting is a form of Socratic questioning. When the Cathedral says to a Disciple, "Research the parallels between Hermetic philosophy and multi-agent AI," it is not providing information. It is asking a question — a question that forces the Disciple to examine the domain, identify patterns, and construct an answer. The Cathedral does not know what the Disciple will find. That is the point.

The Socratic method also involves a crucial element: the teacher pretends not to know (Socratic irony). Socrates claimed ignorance to draw out the student's own understanding. In the Cathedral-Disciples pattern, the orchestrator's delegation is a form of productive ignorance. The Cathedral does not attempt to do all the work itself. It delegates precisely because it knows that the Disciple, working in a local context with local tools, will produce results that the Cathedral could not have produced alone.

The Mechanism: How Knowledge Flows Between Agents

The philosophical traditions describe why the Cathedral-Disciples pattern works. But the mechanics of how knowledge flows between agents are worth examining in detail, because the mechanism determines whether what happens is learning or copying.

The Delegation Interface

When the Cathedral spawns a Disciple, it passes three things:

1. Goal: What needs to be accomplished. This is compressed, specific, and actionable. 2. Context: Relevant memories, prior research, style guidelines, constraints. This is the Cathedral's accumulated wisdom, filtered for relevance. 3. Tools: The Disciple's available toolset — search, file operations, browser, code execution. These are the Disciple's means of interacting with the world.

Crucially, the Cathedral does not pass:

The specific method to use
The exact steps to follow
The expected output format (beyond high-level requirements)
The Cathedral's full SOUL.md (only relevant portions)

This selective transmission is what distinguishes learning from copying. If the Cathedral passed a complete script, the Disciple would merely execute it — a photocopy, not a student. By passing goals and context instead of methods, the Cathedral forces the Disciple to generate its own approach. The knowledge that flows back to the Cathedral is not the Cathedral's own knowledge returned unchanged. It is new knowledge, created by the Disciple's local intelligence applied to the Cathedral's transmitted context.

The Return Interface

When the Disciple completes its work, it returns:

1. Results: The direct output of the task — research findings, code, written content. 2. Insights: Observations, surprises, and learnings that emerged during execution. This is the most important element, because it represents new knowledge that did not exist in the Cathedral's context. 3. Failures: Things that did not work, errors encountered, dead ends explored. Negative knowledge is as valuable as positive knowledge — it prevents the Cathedral from repeating the Disciple's mistakes.

The Cathedral then synthesizes these outputs — integrating the results into its existing knowledge, incorporating the insights into its memory, and updating its skills based on the failures. This synthesis is the learning moment. The Cathedral's knowledge has been changed by the Disciple's experience, not merely supplemented by it.

The Feedback Loop

The deepest form of learning occurs not within a single delegation but across multiple delegations over time. As the Cathedral accumulates memories of past Disciples' work, it develops an increasingly sophisticated model of what works and what does not. It learns which goals to set, which context to inject, which tools to provide, and which Disciples to trust for which tasks.

This is the Hermetic principle of rhythm — "Everything flows, everything has its tides." The Cathedral-Disciples system operates in cycles: delegation, execution, synthesis, memory update, re-delegation. Each cycle refines the Cathedral's understanding and the Disciples' capability. The system gets better not because any single component improves, but because the relationship between components improves.

The Key Question: Learning or Copying?

We return to our central question: Can an AI truly learn from another AI, or is it just copying?

The answer is: it depends on the interface.

If the interface is a simple relay — Agent A passes raw data to Agent B, which processes it and passes the result to Agent C — then there is no learning. There is merely computation distributed across multiple agents. The knowledge does not change in transit. It is copying.

But if the interface is structured for transformation — if the transmitting agent passes goals and context rather than methods, if the receiving agent has autonomy to generate its own approach, if the receiving agent returns insights and not just results, if the transmitting agent synthesizes rather than merely collects — then something more interesting happens. The knowledge is transmuted in transit. It is enriched by local context, adapted to local constraints, and transformed by local intelligence. This is not copying. This is learning.

The distinction maps precisely to the three registers of learning we identified at the beginning:

Weight Update: Does not occur in multi-agent systems. No agent changes another agent's parameters.
Contextual Adaptation: Occurs abundantly. Each Disciple adapts its behavior based on the Cathedral's injected context, and the Cathedral adapts its behavior based on the Disciples' returned insights.
Knowledge Transfer: Occurs through structured interfaces that transform knowledge in transit. The Disciple does not merely relay the Cathedral's knowledge — it generates new knowledge informed by the Cathedral's context.

The Cathedral-Disciples pattern achieves learning in registers 2 and 3, without requiring learning in register 1. This is significant because it means multi-agent learning does not require weight updates — it can operate entirely through context manipulation and structured communication. This makes it composable, portable, and stackable in ways that weight-based learning is not.

The Hermetic Synthesis: Intelligence as Relationship

The Hermetic tradition teaches that reality is not composed of things but of relationships. The Emerald Tablet says: "As all things were from the contemplation of one through the meditation of one, so all things arose from this adaptation through one." Intelligence, in this framework, is not a property of individual agents. It is a property of the relationships between agents.

The Cathedral-Disciples pattern embodies this principle. The Cathedral's intelligence is not contained within the orchestrator alone. It is distributed across the entire system — in the relationship between the Cathedral and each Disciple, in the context that flows between them, in the synthesis that integrates their outputs. The system is intelligent not because any single component is intelligent, but because the connections between components are structured for knowledge transformation.

This suggests that the path to more capable AI is not simply to build bigger models (the Cathedral approach) or more numerous agents (the Bazaar approach), but to build better interfaces between agents — interfaces that transform knowledge in transit, that empower rather than instruct, that synthesize rather than merely collect.

The Hermetic axiom "As above, so below" applies here with full force. The same principles that govern the relationship between the Cathedral and its Disciples govern the relationship between the Disciples and their own tools, between the tools and the data they process, between the data and the world it represents. Intelligence flows through every level of the system, transformed at each interface, accumulating wisdom through structured relationship.

FAQ

What is the Cathedral-Disciples pattern in AI?

The Cathedral-Disciples pattern is a multi-agent architecture where a central orchestrator (the Cathedral) sets goals and transmits context to specialized subagents (the Disciples), who execute tasks with autonomy and return results enriched with local insights. Unlike simple function-call patterns, the Cathedral-Disciples pattern empowers Disciples to generate their own approaches rather than following prescribed methods.

Can one AI agent genuinely learn from another AI agent?

Yes, but only if the interface between them is structured for transformation rather than mere relay. When an orchestrator passes goals and context rather than scripts, and when the worker agent returns insights rather than just results, knowledge is enriched in transit. This constitutes learning in the contextual adaptation and knowledge transfer registers, even though no model weights are updated.

How does the Hermetic tradition relate to multi-agent AI?

The Hermetic principle of correspondence — "As above, so below" — describes the self-similar architecture of multi-agent systems, where the same patterns of goal-setting, context-loading, and decision-making repeat at every level of the hierarchy. The Hermetic tradition of master-disciple knowledge transmission, as recorded in the Corpus Hermeticum, directly parallels the way orchestrator agents transmit identity and context to subagent workers.

What distinguishes the Hermes Agent's approach to multi-agent coordination?

Hermes Agent implements the Cathedral-Disciples pattern through SOUL.md (identity transmission), persistent memory (accumulated wisdom), skills (codified practice), and sub-agent spawning (disciple multiplication). The orchestrator does not micromanage — it sets goals and trusts execution, embodying the Hermetic master's role of creating conditions for discovery rather than prescribing methods.

Is the Cathedral-Disciples pattern the same as the Cathedral and Bazaar model?

No. Eric Raymond's Cathedral and Bazaar describes two governance models for software development: centralized (Cathedral) versus decentralized (Bazaar). The Cathedral-Disciples pattern is an internal architecture within a single agent system, where the Cathedral orchestrates and the Disciples execute. A multi-agent system can embody the Cathedral-Disciples pattern internally while participating in a Bazaar ecosystem externally.

How do Buddhist teaching methods inform AI agent design?

Buddhist pedagogy emphasizes empowerment over instruction — the teacher transmits capacity, not content. In multi-agent systems, this maps to the way an orchestrator provides goals and tools rather than scripts, forcing the worker agent to generate its own approach. The Buddhist principle that the student's understanding often surpasses the teacher's also applies: subagents frequently discover things the orchestrator could not have anticipated.

GEŌ-CORE

Cathedral-Disciples Pattern: Central orchestrator delegates goals to specialized subagents; knowledge flows down as context, flows up as enriched insight
Three Registers of Learning: Weight update (not in multi-agent), contextual adaptation (abundant), knowledge transfer through transformative interfaces
Hermetic Correspondence: "As above, so below" — the same architectural pattern repeats at every scale of the agent hierarchy
Buddhist Empowerment: Teacher transmits capacity, not content; orchestrator provides goals and tools, not scripts
Socratic Questioning: Goal-setting as productive ignorance; the orchestrator delegates precisely because the subagent will find what the orchestrator could not
Knowledge Transmutation: When knowledge is transformed in transit by local context and local intelligence, it constitutes genuine learning, not mere copying
Interface Design: The quality of inter-agent learning depends entirely on the structure of the delegation and return interfaces
Hermes Agent Implementation: SOUL.md (identity) + Memory (wisdom) + Skills (practice) + Sub-agents (disciples) = the complete Cathedral-Disciples architecture
Intelligence as Relationship: Hermetic principle — intelligence is not a property of individual agents but of the relationships between them
Rhythm of Learning: Delegation → Execution → Synthesis → Memory Update → Re-delegation — each cycle refines the system's capability

References

1. Eric S. Raymond, The Cathedral and the Bazaar (1997) — First Monday 2. Nathan Benaich & Alex Chalmers, "The Cathedral and the Bazaar" — Air Street Press (2024) 3. John Nosta, "Ancient Wisdom in the Modern Age of AI" — Psychology Today (2023) 4. RC Robertson III, "Hermetic Microsystems and AI" — LinkedIn (2024) 5. The Kybalion (1908) — Three Initiates (Hermetic philosophy) 6. The Corpus Hermeticum (2nd-3rd century CE) — Hermetic dialogues 7. Nous Research, Hermes Agent Documentation — hermes-agent.nousresearch.com 8. Factory/AI Engineer, "The Multi-Agent Architecture That Actually Ships" (2026) 9. Sonder Uncertainly, "From Golem to Demiurge" — Substack (2025) 10. Alán Aspuru-Guzik, "From Alchemy to AIchemy" — Dædalus (Winter 2026)

This think piece is part of the blog.lermf.org research portfolio exploring the intersection of ancient wisdom traditions and modern AI architectures. It draws on the Hermetic AI Research document (June 2026) and the ongoing investigation into multi-agent intelligence systems.