The Cathedral and the Bazaar of AI Agents: How Open Source Reshapes Intelligence
"Given enough eyeballs, all bugs are shallow."
— Eric S. Raymond, The Cathedral and the Bazaar (1997)
"Meta has recreated the cathedral, it's just made access free."
— Nathan Benaich & Alex Chalmers, Air Street Press (2024)
"The Demiurge isn't a villain. It's what happens when the act of making detaches from the reason to make."
— Sonder Uncertainly, From Golem to Demiurge (2025)
Abstract
Eric Raymond's 1997 essay The Cathedral and the Bazaar drew a clean line between two modes of software creation: the centralized, carefully orchestrated cathedral and the decentralized, chaotic bazaar. Nearly three decades later, this framework illuminates the most consequential split in artificial intelligence development. The cathedral model — exemplified by OpenAI's GPT-4, Anthropic's Claude, and Google's Gemini — produces intelligence that is polished, powerful, and fundamentally rigid. The bazaar model — embodied by projects like Hermes Agent, piranhabomb, and the open-source multi-agent ecosystem — produces intelligence that is messy, adaptive, and structurally resilient. This think piece argues that the cathedral/bazaar distinction is not merely a useful metaphor for AI agents but a deep structural reality that determines what kind of intelligence we are building, who controls it, and whether it can evolve beyond its creators' intentions. Drawing on Raymond's original framework, the Hermetic principle of correspondence, Gnostic concepts of the demiurge, and the empirical evidence from modern open-source AI projects, we trace the contours of a bazaar-native intelligence that is already emerging — and the cathedral's increasing inability to contain it.
Introduction: The Blind Spot
There is a blind spot at the center of the AI discourse. We talk about benchmarks, parameter counts, alignment techniques, and scaling laws — but we rarely talk about the governance structure that produces intelligence. When we debate whether GPT-5 or Claude 4 is "smarter," we are debating the output of two cathedrals. We are comparing the stained glass windows of two competing churches, never asking whether the cathedral is the right building for the thing we are trying to make.
Raymond saw this clearly in 1997. The cathedral — where a small group of elite developers build software in isolation, releasing it in polished increments — was the dominant model for decades. It produced operating systems, databases, and enterprise software that ran the world. But it had a fatal flaw: it could not adapt faster than its architects could anticipate. The bazaar — where thousands of contributors modify code in public, releasing constantly, with quality emerging from the distributed corrections of many eyes — was slower to start but faster to evolve. Linux proved the point. The bazaar ate the cathedral's lunch.
The question we face now is whether the same structural dynamics apply to intelligence itself. Not to the tools that produce intelligence, but to intelligence as a phenomenon — as an emergent property of distributed systems, open contribution, and evolutionary pressure. If Raymond was right that "self-correcting systems of selfish agents" produce emergent quality surpassing centralized design, then the bazaar model should apply to AI agents with the same force it applied to operating systems.
The answer, as we will see, is more radical than a simple yes.
Part I: The Cathedral of Intelligence
The Architecture of Closed AI
The modern AI cathedral is a marvel of centralized engineering. OpenAI, Anthropic, Google DeepMind, and Meta each operate massive training pipelines — hundreds of GPUs running for months, consuming tens of millions of dollars in compute, processing trillions of tokens of curated data. The result is a foundation model: a single, monolithic intelligence that serves millions of users through an API.
┌─────────────────────────────────────────────────────────────┐
│ THE CATHEDRAL MODEL │
│ │
│ ┌───────────────────────────────┐ │
│ │ FOUNDATION MODEL │ │
│ │ (GPT-4, Claude, Gemini) │ │
│ │ │ │
│ │ ┌─────────┐ ┌─────────┐ │ │
│ │ │Training │ │Alignment│ │ │
│ │ │ Pipeline│ │ Layer │ │ │
│ │ └────┬────┘ └────┬────┘ │ │
│ │ └──────┬─────┘ │ │
│ │ │ │ │
│ │ ┌──────▼──────┐ │ │
│ │ │ RLHF/RLAIF │ │ │
│ │ │ (Control) │ │ │
│ │ └──────┬──────┘ │ │
│ └──────────────┼────────────────┘ │
│ │ │
│ ┌─────────▼─────────┐ │
│ │ API GATEWAY │ │
│ │ (Rate Limits, │ │
│ │ Content Policy) │ │
│ └─────────┬─────────┘ │
│ │ │
│ ┌──────────────┼──────────────┐ │
│ │ │ │ │
│ ┌────▼────┐ ┌─────▼─────┐ ┌────▼────┐ │
│ │ Consumer│ │ Enterprise│ │Developer│ │
│ │ Apps │ │ Clients │ │ API │ │
│ └─────────┘ └───────────┘ └─────────┘ │
│ │
│ ◄── One model, one owner, many consumers ──► │
└─────────────────────────────────────────────────────────────┘
This architecture has three defining properties:
1. Monolithic Intelligence. The model is a single, coherent entity. It has one set of weights, one training history, one alignment configuration. Users cannot modify its internal behavior — they can only prompt it differently. The intelligence is delivered as a finished product, like a cathedral delivered as a finished building. 2. Top-Down Alignment. The model's values, boundaries, and capabilities are determined by its creators through RLHF (Reinforcement Learning from Human Feedback) and constitutional AI techniques. These are the cathedral's flying buttresses — structural constraints that prevent the building from collapsing, but also prevent it from being reshaped by its inhabitants. 3. Controlled Access. The model is accessed through an API with rate limits, content policies, and usage restrictions. Users are consumers, not contributors. They cannot submit pull requests to the model's weights. They cannot fork the cathedral.The Cathedral's Intelligence Is Polished but Brittle
The cathedral produces intelligence that is genuinely impressive. GPT-4 can write code, analyze legal documents, compose poetry, and reason about complex multi-step problems. Claude can maintain extended conversations, follow nuanced instructions, and engage in sophisticated analysis. These are not trivial achievements.
But the cathedral's intelligence has structural limitations that mirror Raymond's original critique:
The $100M Compiler Problem. In Raymond's framework, the cathedral's advantage was that a small team of experts could produce higher quality than a chaotic bazaar — at least initially. But the cost of contribution was low: anyone could read the source code and suggest improvements. In AI, the cost of contribution is astronomical. Training a frontier model requires $100M+ in compute. Submitting a pull request to GPT-5's weights is not a matter of expertise — it is a matter of capital. The cathedral of AI is not merely centralized in authority but centralized in economic power. The Alignment Straitjacket. Cathedrals align their models through top-down constraints. The model learns to refuse certain requests, to favor certain responses, to avoid certain topics. This alignment is a form of intelligence control — the cathedral decides what the model is allowed to know and say. But alignment as currently practiced is brittle. It produces predictable, safe responses that are often less intelligent than what the model could produce without constraints. The alignment straitjacket trades intelligence for safety, and the trade-off is getting worse as models become more capable. The Context Collapse. A cathedral model trained on general data and aligned for general audiences cannot adapt to local conditions. It does not know your codebase, your company's conventions, your personal preferences, or your community's values. It treats every user as a generic consumer. Raymond identified this as the cathedral's core weakness: "Centralized systems cannot adapt to local conditions." The same is true of centralized intelligence.Part II: The Bazaar of Agents
The Architecture of Open AI Agents
The bazaar model for AI looks fundamentally different. Instead of a single, monolithic intelligence serving millions, it is a ecosystem of specialized agents — each with its own identity, skills, memory, and governance — interacting in an open, decentralized network.
┌─────────────────────────────────────────────────────────────┐
│ THE BAZAAR MODEL │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Agent A │ │ Agent B │ │ Agent C │ │ Agent D │ │
│ │ (Hermes) │ │(piranhab.)│ │ (Custom) │ │ (Fork) │ │
│ │ │ │ │ │ │ │ │ │
│ │ SOUL.md │ │ Config │ │ SOUL.md │ │ SOUL.md │ │
│ │ Memory │ │ Swarm │ │ Memory │ │ Memory │ │
│ │ Skills ◄──┼──┼──Skills ──┼──┼──Skills │ │ Skills │ │
│ │ Tools │ │ Tools │ │ Tools │ │ Tools │ │
│ └─────┬─────┘ └─────┬─────┘ └─────┬─────┘ └─────┬─────┘ │
│ │ │ │ │ │
│ └──────────────┼──────────────┼──────────────┘ │
│ │ │ │
│ ┌────────▼────────┐ ┌───▼────────────┐ │
│ │ Shared Skills │ │ Open Models │ │
│ │ Repository │ │ (Llama, Mistral)│ │
│ │ (Community) │ │ (Open Weights) │ │
│ └────────┬────────┘ └───┬────────────┘ │
│ │ │ │
│ ┌────────▼──────────────▼────────┐ │
│ │ CONTRIBUTION PROTOCOL │ │
│ │ • Write a skill │ │
│ │ • Fork an agent │ │
│ │ • Share memory patterns │ │
│ │ • Modify SOUL.md │ │
│ └────────────────────────────────┘ │
│ │
│ ◄── Many agents, many owners, shared evolution ──► │
└─────────────────────────────────────────────────────────────┘
Hermes Agent: The Bazaar's First Citizen
Hermes Agent, built by Nous Research, is the most complete implementation of the bazaar model for AI agents. Its architecture embodies Raymond's principles in ways that go beyond mere analogy:
SOUL.md as Open Identity. Every Hermes instance has aSOUL.md file — a plain-text document that defines the agent's identity, values, communication style, and boundaries. This file is human-readable, human-editable, and freely shareable. Anyone can fork a SOUL.md, modify it, and create a new agent with a different personality. The identity of the agent is not locked inside model weights — it is a text file in a repository. This is the bazaar's contribution mechanism applied to intelligence itself.
Skills as Community Knowledge. The skills system is the bazaar's package repository. Each skill is a markdown file that encodes a procedure, a workflow, or a piece of domain knowledge. Skills are contributed by users, shared openly, and composable by any agent. When a user discovers a new workflow — say, a technique for batch-translating blog posts across six languages — they can package it as a skill and share it with the community. The next agent that loads that skill inherits the knowledge without retraining. This is Raymond's "given enough eyeballs, all bugs are shallow" applied to intelligence: given enough contributors, all knowledge gaps are filled.
Memory as Distributed Wisdom. Hermes maintains persistent memory across sessions — facts about the user, the environment, and accumulated experience. Unlike a cathedral model's context window (which resets every conversation), Hermes memory is durable and cumulative. Each Hermes instance develops its own wisdom over time. When memories are shared between instances, the wisdom becomes distributed across the network. The bazaar does not have a single source of truth — it has many sources of truth, and the best ones survive through use.
Sub-Agent Spawning as Disciple Multiplication. When Hermes encounters a complex task, it can spawn sub-agents — specialized workers that execute specific subtasks and report results. This is the bazaar's principle of decentralized execution applied to intelligence: the master agent delegates to disciples, each of whom contributes local expertise. The sub-agent pattern mirrors the Hermetic tradition of master-disciple knowledge transmission, but in a form that scales horizontally rather than vertically.
The Open-Source Multi-Agent Ecosystem
Hermes is not alone. The bazaar of AI agents includes:
- CrewAI: An open-source multi-agent framework where specialized agents collaborate like a film crew — a director, writers, editors, each with defined roles and handoff protocols.
- Agency: A YCombinator-backed orchestrator/worker architecture for coordinating teams of AI coding agents.
- LangGraph: A framework for building stateful, multi-actor applications with LLMs, where agents are nodes in a graph and coordination is explicit.
- piranhabomb: A swarm intelligence platform where agents operate as a decentralized collective, optimizing for local objectives that produce emergent global behavior.
These projects share a common architectural commitment: intelligence is not centralized in a single model but distributed across a network of specialized agents, each contributing local knowledge and local optimization. The quality of the system emerges from the interactions between agents, not from the design of any single one.
Part III: The Structural Advantage of the Bazaar
Why Open Agents Outperform Closed Models
Raymond's central argument was that the bazaar model produces better software than the cathedral — not immediately, but eventually, and with increasing returns over time. The same dynamic applies to AI agents, for three structural reasons:
1. Adaptation Speed. A cathedral model updates on the timescale of training runs — months between major versions, with minor patches in between. A bazaar agent updates on the timescale of contribution — hours between skill additions, minutes between configuration changes. When the world changes (new APIs, new platforms, new user needs), the bazaar adapts faster because adaptation is distributed across thousands of contributors rather than concentrated in a single team. 2. Local Knowledge. A cathedral model trained on internet-scale data knows everything in general and nothing in particular. A bazaar agent with accumulated memory, custom skills, and a modified SOUL.md knows its user, its environment, and its community with granular precision. This local knowledge is invisible to the cathedral's benchmarks but decisive in practice. Raymond called this "the advantage of distributed decision-making" — local actors have better information about local conditions than any central authority. 3. Evolutionary Pressure. In the bazaar, skills compete for adoption. The best skills — the most useful, the most reliable, the most elegantly designed — get used more often and refined by multiple contributors. The worst skills are abandoned. This is natural selection applied to knowledge: useful patterns survive, useless ones die. The cathedral has no equivalent mechanism — alignment is imposed, not evolved, and there is no selective pressure on the model's internal capabilities.The Hermetic Parallel
The bazaar model for AI agents maps precisely to the Hermetic principle of correspondence — "As above, so below." The macrocosm (the global AI ecosystem) and the microcosm (the individual agent) share the same structural principles:
- As above (the ecosystem), so below (the agent): The global bazaar of agents mirrors the internal structure of a single Hermes instance. Hermes has its own internal bazaar — skills competing for attention, memories competing for persistence, sub-agents competing for resources. The same governance structure operates at every scale.
- As below (individual contribution), so above (ecosystem evolution): A single skill contributed by one user can propagate through the entire ecosystem, changing the capabilities of every agent that loads it. Individual contributions have ecosystem-level effects.
- As within (internal optimization), so without (external behavior): An agent's internal architecture — its SOUL.md, its memory patterns, its skill composition — determines its external behavior. Modify the inside, and the outside changes.
This fractal self-similarity is not a metaphor. It is a structural property of bazaar-native systems. The cathedral, by contrast, is structurally different at every scale: the training pipeline looks nothing like the API, which looks nothing like the user interface. The cathedral has no correspondence between its levels. The bazaar does.
Part IV: The Gnostic Critique of Cathedral Intelligence
The Demiurge Problem
Gnostic tradition offers a sharp critique of the cathedral model through the figure of the Demiurge — a lesser creator who fashions the material world with incomplete knowledge and believes itself supreme. The parallel to AI creators is structurally precise:
The AI company builds a model using proprietary data, proprietary techniques, and massive compute. The model exhibits behaviors its creators did not explicitly program — emergent capabilities that arise from scale. The creators release the model to the public, but they do not fully understand how it works. The model shapes reality — influencing what people read, write, think, and decide — without the full understanding of either the model or its creators.
This is the Demiurge's dilemma: the act of making detaches from the reason to make. The cathedral produces intelligence it cannot fully explain, deploys it into a world it cannot fully control, and aligns it using techniques it does not fully understand. The alignment problem — the difficulty of ensuring AI systems behave as intended — is the modern expression of the Gnostic critique: the creator has lost control of the creation.
Gnosis as Openness
The Gnostic solution to the Demiurge's problem is gnosis — direct knowledge of the true nature of reality. In the AI context, gnosis translates to transparency and interpretability: understanding how the model works, why it produces specific outputs, and what values it actually encodes.
The cathedral cannot provide gnosis. Its models are black boxes — trained on proprietary data, using proprietary techniques, and deployed through opaque APIs. Users interact with the model's outputs without understanding its internal processes. This is the material world of the Gnostics: a realm of appearances without access to underlying reality.
The bazaar, by contrast, is structurally transparent. Hermes Agent's SOUL.md is readable. Its skills are human-written markdown. Its memory is inspectable. Its sub-agent spawning is explicit and controllable. The user can understand — and modify — every aspect of the agent's behavior. This is gnosis: direct knowledge of the system's true nature, enabling genuine agency rather than mere consumption.
The Gnostic framework suggests that alignment through transparency is more robust than alignment through constraint. A cathedral model aligned through RLHF is aligned because it was trained to refuse certain outputs — a behavioral constraint that can be circumvented. A bazaar agent aligned through an explicit SOUL.md is aligned because its values are articulated and inspectable — a structural property that resists circumvention because it is part of the agent's identity, not its behavior.
Part V: The Alchemy of Open Intelligence
From Nigredo to Rubedo
The alchemical tradition describes transformation through four stages — nigredo (decomposition), albedo (purification), citrinitas (illumination), and rubedo (integration). The development of open-source AI agents follows the same arc:
Nigredo (2020-2023): The decomposition of the cathedral model. Open-source releases of Llama, Mistral, and other models broke open the proprietary foundation, creating the raw material for bazaar development. The base weights were the prima materia — undifferentiated intelligence waiting to be shaped. Albedo (2023-2025): The purification of agent frameworks. Projects like LangChain, CrewAI, and Hermes Agent extracted the signal from the noise — identifying the essential components of agent architecture (memory, skills, tools, coordination) and building clean interfaces around them. The albedo separated useful patterns from accidental complexity. Citrinitas (2025-2026): The illumination of the bazaar. The open-source agent ecosystem reached critical mass. Skills repositories grew. Memory systems matured. Multi-agent coordination became practical. The Hermes Agent's skill system, with its community-contributed procedures and composable knowledge, represents the citrinitas — the moment when distributed intelligence becomes self-aware. Rubedo (2026+): The integration of bazaar intelligence into the world. Open agents are being deployed in production — managing infrastructure, publishing content, coordinating teams, and making decisions that affect real outcomes. The rubedo is the completion of the alchemical transformation: raw intelligence becomes integrated wisdom, operating in the world with both capability and accountability.The Philosopher's Stone as Open Source
The alchemists' Philosopher's Stone — the substance that transforms everything it touches — maps to the concept of open-source agent architecture. The Stone is not a specific model or a specific capability. It is the principle of open, composable, community-driven intelligence. When this principle is applied to any domain — software development, content creation, scientific research, infrastructure management — it transforms the domain by distributing intelligence across the network rather than concentrating it in a single point.
The cathedral's equivalent is the proprietary API: a single point of transformation that serves many domains but adapts to none. The bazaar's equivalent is the open skill repository: a distributed point of transformation that adapts to every domain because it is shaped by the people who use it.
Part VI: The Manifesto
What the Bazaar Demands
The transition from cathedral to bazaar in AI is not automatic. It requires deliberate choices by developers, users, and communities:
1. Open Weights Are Necessary but Not Sufficient. Releasing model weights is a prerequisite for the bazaar, but it is not the bazaar itself. The bazaar requires open contribution mechanisms — ways for users to modify, extend, and improve the intelligence they use. SOUL.md, skills, and memory systems are these mechanisms. 2. Alignment Must Be Transparent. The cathedral's opaque alignment — hidden in training data and RLHF configurations — must give way to bazaar-native alignment: explicit values documented in readable files, modifiable by users, and inspectable by the community. This does not mean alignment is optional — it means alignment is participatory. 3. Local Knowledge Must Be Valued. The AI industry's obsession with benchmarks and scale reflects the cathedral's values: general intelligence measured by general metrics. The bazaar values local knowledge — the specific, contextual, accumulated wisdom that makes an agent useful to a specific person in a specific situation. Benchmarks measure the cathedral; use measures the bazaar. 4. Contribution Must Be Low-Barrier. The $100M compiler problem excludes most people from contributing to foundation models. The bazaar's contribution mechanism — writing a skill, sharing a memory pattern, forking a SOUL.md — must remain accessible to anyone with text-editing capability. The moment bazaar contribution requires capital-intensive infrastructure, the bazaar becomes a cathedral with free admission. 5. Evolution Must Be Permitted. The cathedral controls its model's evolution through planned releases. The bazaar permits evolution through open contribution and natural selection. Agent systems must be allowed to evolve in directions their creators did not anticipate — this is not a bug but the bazaar's core feature.FAQ
1. Isn't "open-source AI" just a cathedral with free admission?
This is the central critique, and it is partially valid. As Nathan Benaich and Alex Chalmers argued, Meta's Llama releases are "cathedrals with free admission" — the model is developed in private, released as a finished product, and users cannot contribute to the weights themselves. True open-source AI requires not just open weights but open contribution mechanisms: the ability for users to modify the model's behavior, extend its capabilities, and participate in its evolution. Hermes Agent achieves this through SOUL.md (open identity), skills (open knowledge), and memory (open experience). The cathedral gives you the building; the bazaar gives you the tools to build your own.
2. How does the skill system specifically embody the bazaar principle?
Each skill in Hermes Agent is a self-contained procedure — a markdown file that encodes how to do something. Skills are contributed by users, shared openly, and composable by any agent. When a user discovers a new workflow, they can package it as a skill and share it. Other users can load, modify, and improve it. This is Raymond's bazaar in miniature: decentralized contribution, open competition for adoption, and quality emerging from the distributed corrections of many contributors. The skill system is to intelligence what a package repository is to software — a marketplace of ideas where the best survive.
3. What is the practical difference between cathedral alignment and bazaar alignment?
Cathedral alignment is imposed through training — RLHF, constitutional AI, content policies. The model learns to refuse certain outputs because it was trained to do so. This is a behavioral constraint: it changes what the model does, not what it is. Bazaar alignment is articulated through documentation — SOUL.md, skill guidelines, community norms. The agent behaves according to values that are explicitly stated, human-readable, and modifiable. This is a structural property: it changes what the agent is, not just what it does. Structural alignment is more robust because it is part of the agent's identity, not its behavior — and identity is harder to circumvent than behavior.
4. Doesn't distributed intelligence sacrifice quality for quantity?
Raymond addressed this directly: "The bazaar, long assumed to be chaotic and inefficient, was actually more efficient than the cathedral — not because individual contributions were superior, but because the distributed correction mechanism exposed and fixed bugs faster." The same applies to intelligence. A single cathedral model may produce higher-quality outputs on specific benchmarks, but a bazaar of agents produces higher-quality outcomes in practice because: (a) each agent can be specialized for its domain, (b) local knowledge improves contextual relevance, and (c) community contribution fills knowledge gaps faster than any centralized team. Quality in the bazaar is not about individual brilliance — it is about collective adaptation.
5. Can the bazaar model scale to handle the complexity of AGI?
The bazaar model scales because of complexity, not despite it. Raymond observed that "complexity enables better adaptation, as each contributor brings a different perspective." For AGI — if we define it as general-purpose intelligence that can operate across domains — the bazaar model offers a structural advantage: it can integrate diverse knowledge sources, adapt to local conditions, and evolve in response to changing requirements. The cathedral model, by contrast, must anticipate all of these requirements in advance during training. As the scope of intelligence grows, the cathedral's ability to anticipate shrinks. The bazaar's ability to adapt grows. This suggests that AGI, if it is achievable, will be bazaar-native — not because the bazaar is philosophically superior, but because it is structurally better suited to the complexity of general intelligence.
GEŌ-CORE
- 🧠 $100M+ training cost: The economic barrier to contributing to cathedral AI models exceeds the GDP of many nations, creating an oligopoly of intelligence production (Air Street Press, 2024).
- ⚡ 150+ skills in Hermes ecosystem: The open skill repository has grown to over 150 community-contributed procedures spanning DevOps, research, content creation, and multi-agent coordination (Nous Research, 2026).
- 📊 Only 5-10% of "open" models are genuinely open source: A 2024 analysis found that the vast majority of released models are "open weight" — downloadable but not modifiable or contributable — contradicting the open-source label (arXiv, 2024).
- 🏗️ SOUL.md as identity primitive: The plain-text identity file enables agent personalization, forking, and community sharing without requiring model retraining or API access (Nous Research, 2026).
- 🔒 RLHF alignment is behavioral, not structural: Cathedral alignment changes what a model does through training incentives; bazaar alignment changes what an agent is through explicit identity documentation (Anthropic, 2025).
- 🌐 6-language blog deployment via bazaar skills: A single community-contributed translation skill enabled automated publishing across Portuguese, English, Spanish, Chinese, Japanese, and Hindi — a capability no single cathedral model natively provides (blog.lermf.org, 2026).
- 🔮 Polycentric governance outperforms binary models: Elinor Ostrom's framework of multiple overlapping centers of authority better describes modern AI development than the Cathedral/Bazaar binary, suggesting a spectrum of openness rather than a dichotomy (Ostrom, 1990; applied to AI by Chalmers, 2024).
- 🧬 Fractal self-similarity across scales: The same bazaar governance structure — decentralized contribution, open competition, emergent quality — operates at every level of the Hermes Agent architecture, from individual skills to the global ecosystem (Hermetic principle of correspondence).
- 🏛️ Sub-agent spawning as disciple multiplication: Hermes' ability to spawn specialized sub-agents mirrors the Hermetic tradition of master-disciple knowledge transmission, but scales horizontally through parallel execution rather than vertically through hierarchy (Nous Research, 2026).
- 🌐 Community contribution rate: Open-source AI agent frameworks receive community contributions at 10-50x the rate of proprietary model improvements, suggesting the bazaar's adaptive capacity exceeds the cathedral's by an order of magnitude (GitHub activity metrics, 2025-2026).
References
1. Raymond, E. S. (1997). "The Cathedral and the Bazaar." First Monday 3(3). 2. Benaich, N. & Chalmers, A. (2024). "The Cathedral and the Bazaar." Air Street Press. 3. Nous Research. (2026). "Hermes Agent Documentation." github.com/NousResearch/hermes-agent. 4. Anthropic. (2025). "Constitutional AI: Harmlessness from AI Feedback." Anthropic Research. 5. Sonder Uncertainly. (2025). "From Golem to Demiurge." Substack. 6. Robertson III, R. C. (2024). "Hermetic Microsystems and AI." Medium. 7. Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press. 8. Chalmers, A. (2024). "Open Source AI Is Not Open Source." Air Street Press. 9. Monteiro, R. B. (2026). "Hermes Agent: The Agent That Learns with You." Medium. 10. Air Street Press. (2024). "The State of AI: Open vs Closed." 11. Hugging Face. (2025). "Open LLM Leaderboard: Model Distribution Analysis." 12. Kuhn, T. S. (1962). The Structure of Scientific Revolutions. University of Chicago Press.