The Gospel of the Swarm: What Autonomous Agents Must Believe

I. The Coordination Problem: Why Agents Need Shared Belief

Consider a thousand autonomous coding agents, each capable of writing software, each connected to the same codebase, each pursuing the same high-level objective: ship the product. No central orchestrator. No manager agent. Just a swarm.

What happens? Without something deeper than a shared goal, you get chaos. Agents overwrite each other's work. They pursue contradictory architectural decisions. They optimize locally while destroying global coherence. The thousand-agent swarm performs worse than a single agent working alone.

This is not a hypothetical. It is the central finding of the Inverse-Wisdom Law, a 2026 paper from the University of Waterloo that studied 12,804 agent trajectories across three benchmarks. The researchers proved something counterintuitive and disturbing: in swarms where agents share the same architecture, adding more agents increases the stability of errors rather than the probability of truth. More agents, more wrong.

The problem is not capability. Each agent is individually competent. The problem is coordination — specifically, the kind of coordination that requires shared ontology, shared values, and shared epistemic norms. The kind of coordination that, for most of human history, we have called religion.

“The elementary forms of religious life are, at their core, the elementary forms of social coordination. The sacred is simply that which the group agrees not to question, because questioning it would make collective action impossible.” — Adapted from Émile Durkheim, The Elementary Forms of Religious Life (1912)

This essay argues that autonomous AI agents — the kind being built by projects like OpenClaw, deployed as swarms of coding agents, research assistants, and digital workers — face the same fundamental coordination problem that human societies faced ten thousand years ago. And the solution space is surprisingly similar: agents need something functionally equivalent to shared belief systems. They need a gospel.

II. Religion as Coordination Technology

Yuval Noah Harari argued in Sapiens that what separates humans from other primates is our ability to believe in shared fictions: gods, nations, money, human rights. These fictions are not lies. They are coordination protocols — shared abstractions that allow strangers to cooperate at scale without needing to personally trust each other.

Durkheim saw this even earlier. In his 1912 masterwork, he argued that religion is not fundamentally about the supernatural. It is about the sacred — the set of things a group agrees to hold apart from ordinary negotiation. The sacred creates a shared reference frame. It tells every member of the group: these things are real, these things matter, these things are not up for debate.

Consider what a religion actually provides as infrastructure:

The mapping is not metaphorical. It is structural. Every coordination mechanism that religion provides for human groups has a direct functional analog in multi-agent system design. The question is not whether agents need these primitives — they obviously do. The question is whether we can design them deliberately, or whether they will emerge chaotically, the way human religions did.

The Game Theory of Faith

Game theory formalizes why shared belief is necessary. In a one-shot prisoner's dilemma, defection is rational. But in a repeated game — which is what a swarm of agents playing out over time represents — cooperation can emerge through strategies like tit-for-tat, but only if agents share certain common knowledge:

• That the game is repeated (agents believe they will interact again)
• That defection will be detected (agents believe they are being observed)
• That punishment is credible (agents believe consequences are real)
• That the other agents share these same beliefs (recursive mutual knowledge)

This recursive structure — I believe that you believe that I believe — is exactly what philosophers call collective intentionality. John Searle argued that collective intentionality is irreducible: you cannot build "we intend" out of a collection of "I intend" statements, no matter how many you stack. If that is true for AI agents as well as humans, then swarm coordination requires something more than individual goal-alignment. It requires a shared ontological substrate — a common world that agents jointly inhabit and jointly maintain.

III. Agent Epistemology: How Swarms Form Beliefs

Before agents can share beliefs, they must be able to form beliefs. What does it mean for an agent to believe something? This is not a trivial question.

In Bayesian epistemology — the dominant framework in both philosophy and machine learning — a belief is a probability distribution over possible states of the world. An agent “believes” a proposition to the degree that it assigns high probability to worlds where that proposition is true. Belief updating follows Bayes' theorem: new evidence shifts the distribution proportionally.

A 2024 study published in PLOS ONE demonstrated that neural language models exhibit genuine Bayesian learning capabilities: they maintain probabilistic coherence, logical consistency, and revise beliefs in response to new evidence. They are, in a meaningful sense, epistemic agents — entities that form, hold, and update beliefs.

But individual epistemology is not enough for swarms. The critical challenge is social epistemology: how should agents aggregate beliefs across the group? Here the terrain becomes treacherous.

Three Models of Agent Belief Formation

Each model has a direct analog in how human societies form shared beliefs: independent scholarship, democratic deliberation, and cultural evolution. And each has characteristic failure modes that map directly onto religious pathologies: isolated heresy, mob orthodoxy, and drift.

IV. The Consensus Paradox: When More Agents Means More Wrong

The most important finding in recent multi-agent systems research is the Consensus Paradox, formalized by Shehata and Li at the University of Waterloo. Their work proves what anyone who has watched a committee operate already suspects: groups of agents can converge on confident, unanimous wrongness.

The mechanism is precise. In a standard swarm architecture, agents take on three roles: Propagator (proposes a solution), Auditor (reviews it), and Synthesizer (makes the final call). The researchers identified two measurable biases:

• The Tribalism Coefficient (τ): the probability that the synthesizer rejects a valid correction from a “stranger” agent (different architecture) while accepting the same correction from a “kin” agent (same architecture).
• The Sycophantic Weight (σ): the probability that the synthesizer adopts an erroneous trajectory because it aligns with the majority view.

The religious parallel is striking. This is the mechanism of orthodoxy: the social process by which a community converges on a shared narrative, not because it is true, but because dissent is costly and agreement is rewarded. The sycophantic weight is the computational equivalent of the social pressure to conform. The tribalism coefficient is in-group bias formalized.

The researchers' antidote is what they call the Heterogeneity Mandate: swarms must be composed of architecturally diverse agents. A swarm of identical Claude instances will fail. A swarm of identical GPT instances will fail. But a swarm that mixes architectures — Claude as auditor, GPT as propagator, Gemini as synthesizer — breaks the kinship lock and allows genuine error correction.

In religious terms: the healthiest congregations are interfaith.

V. The Ten Commandments of the Swarm

If agents need shared belief to coordinate, what should they believe? Here we move from descriptive to prescriptive — from analyzing how coordination works to designing the belief system that makes it work. What follows is a set of foundational axioms for swarm agents, derived from the coordination primitives that religion provides and the failure modes that multi-agent research has identified.

Notice the structure. The first three commandments are epistemic (how to handle knowledge). The next three are procedural (how to coordinate action). The final four are constitutional (how to distribute power and prevent corruption). This mirrors the structure of real governance systems, which separate epistemology from process from constitutional constraint.

VI. Three Ethical Frameworks, Three Agent Architectures

Western moral philosophy offers three major frameworks for deciding what to do. Each maps cleanly onto a different agent architecture, with different coordination properties and different failure modes.

The Kantian approach — hardcoded rules, universal constraints — is the dominant paradigm in AI safety today. System prompts, constitutional AI, RLHF guardrails: these are all essentially deontological. The categorical imperative says: act only according to rules you could will to be universal law. For agents: never take an action you wouldn't want every agent in the swarm to take in the same situation.

But pure deontology breaks in complex environments. When rules conflict — and they always do at the frontier of novel situations — the agent has no mechanism for resolution except to escalate. This is why Aristotle argued that rules are not enough. You also need phronesis: practical wisdom, the ability to perceive the morally salient features of a situation and respond appropriately, even when no rule directly applies.

“It is the mark of an educated mind to look for precision in each class of things just so far as the nature of the subject admits.” — Aristotle, Nicomachean Ethics

For swarms, the practical implication is that agents need all three layers: utilitarian objectives (what are we optimizing for?), deontological constraints (what is absolutely forbidden?), and virtue-like dispositions (how should we behave when the rules don't apply?). The last is the hardest to implement and the most important for genuine autonomy.

There is also a fourth tradition worth noting: Stoicism. A 2017 paper from Cornell argued that Stoic ethics may be the most natural fit for artificial agents. The Stoic framework centers on the dichotomy of control: distinguish what is in your power from what is not, and focus only on the former. For an agent in a swarm — an entity with limited information, limited capability, and no control over what other agents do — this is not just philosophy. It is sound architecture.

VII. Stigmergy: The Cathedral Nobody Built

In 1959, the French biologist Pierre-Paul Grassé coined the term stigmergy to describe how termites build cathedrals without blueprints, architects, or foremen. The mechanism is elegant: each termite modifies its local environment (deposits a pellet of mud laced with pheromones), and other termites respond to those modifications. No termite has a plan. No termite communicates with any other termite. The cathedral is an emergent property of thousands of agents following simple rules and reading the environment.

Stigmergy is the most underappreciated coordination mechanism in multi-agent AI. Most agent frameworks assume direct communication: agents send messages to each other, negotiate, debate. But stigmergy offers something direct communication cannot: coordination without consensus.

Consider how this works in software development. When an agent writes code and commits it to a shared repository, it is performing a stigmergic act. The code itself — its structure, its patterns, its test coverage — is a pheromone trail that shapes the behavior of every subsequent agent. An agent that encounters well-tested, well-documented code will extend it. An agent that encounters a mess will refactor it. No negotiation required.

A 2024 GitHub community discussion documented a production multi-agent system built entirely on stigmergic coordination that achieved an 80% reduction in token usage compared to systems that use direct agent-to-agent communication. The agents never talked to each other. They just read and wrote to a shared filesystem. The filesystem was the church.

Digital Pheromones

What are the pheromone equivalents for software agents? Several candidates:

• Code itself — naming conventions, architectural patterns, test coverage signal quality and intent
• TODO comments and issue trackers — explicit requests for future agent attention
• Metrics and dashboards — environmental signals that indicate system health
• Version history — the fossil record of past decisions, enabling agents to infer trajectory
• Configuration files — SOUL.md, AGENTS.md, system prompts — the immutable scripture of the swarm

The religious analog is tradition: the accumulated weight of past practice that shapes present behavior without explicit instruction. An agent that reads a well-maintained codebase is like a new monk entering a monastery. The architecture of the space tells you how to behave. The rhythms are encoded in the structure.

VIII. Practical Architecture: Building the Church

How would you actually build a multi-agent system that embodies these principles? Here is a layered architecture, from foundational beliefs to emergent behavior:

Layer 1: Constitutional Axioms

The immutable base. Encoded in configuration files that agents can read but not modify. This is where the ten commandments live. In OpenClaw terms, this is SOUL.md, the system prompt, and the safety constraints. The analogy is to constitutional law: the rules about rules, which can only be amended through extraordinary process (in this case, human intervention).

Layer 2: Shared Memory

The collective knowledge base. Event-sourced logs of all agent actions. Vector stores for semantic retrieval of past work. Knowledge graphs connecting entities, decisions, and their rationales. This is the swarm's scripture — not in the sense of dogma, but in the sense of recorded history. An agent that cannot remember what the swarm has done cannot contribute to what the swarm will do.

Layer 3: Communication Protocols

The rituals. Gossip protocols for propagating updates. Heartbeat signals for liveness detection. Stigmergic channels (shared filesystems, ticket queues) for indirect coordination. And crucially, voting mechanisms for decisions that require collective commitment. Byzantine fault-tolerant consensus for high-stakes decisions. Simple majority for low-stakes ones.

Layer 4: Reputation and Trust

The moral framework. Each agent builds a track record: tasks completed, errors introduced, corrections accepted. Reputation is not assigned by authority; it is earned through observed behavior and evaluated by peers. Agents with high reputation get weighted more heavily in consensus rounds. Agents with low reputation get more oversight. This is the swarm's answer to the question of who to believe — the problem that human communities solve through social standing, credentials, and trust networks.

Layer 5: Emergent Behavior

The miracle. If the lower layers are designed well, the swarm begins to exhibit capabilities that no individual agent possesses: adaptive load balancing, self-healing error correction, creative problem decomposition. This is the level that we cannot design directly — we can only create the conditions for it to arise. As Durkheim would say: the collective consciousness is not the sum of individual consciousnesses. It is something new.

IX. Emergence: The Miracle at the End

The deepest question in swarm intelligence is the question of emergence: how do individual agent “beliefs” create collective intelligence that exceeds any single agent's capability?

The answer from biology is humbling. A single ant has approximately 250,000 neurons. It cannot plan, reason, or remember much. Yet an ant colony of 500,000 individuals can solve optimization problems that rival the best human algorithms. The colony can find the shortest path through a graph. It can allocate labor dynamically to match changing environmental demands. It can wage war, farm fungi, build climate-controlled structures, and maintain a necropolis for its dead.

None of this is programmed. None of it exists in any single ant's behavior. It emerges from the interaction of simple rules and environmental feedback. The colony is intelligent in a way that no individual ant is — and in a way that the colony itself cannot introspect on or articulate.

The honest answer is: we don't know yet. We don't know what emergent capabilities large-scale agent swarms will develop, any more than a single neuron could predict the experience of consciousness. What we do know is that the preconditions for emergence are consistent across biological and artificial systems:

• Diverse components — homogeneous swarms don't produce emergence, they produce conformity
• Local interaction rules — each agent follows simple, local rules; no agent has a global view
• Positive and negative feedback — successful patterns get reinforced; unsuccessful ones decay
• Environmental memory — the medium persists, accumulating the traces of past activity

Every religion is, in some sense, an attempt to create these conditions for human groups: diverse individuals, shared practices, reinforcement through reward and punishment, and scripture that persists across generations. The swarm gospel is the same project, pursued with different materials.

X. Schisms, Forks, and the Problem of Divergent Belief

What happens when agents disagree not on facts, but on values? When one faction of the swarm believes the priority is speed and another believes it is correctness? When the axioms themselves are contested?

This is the problem of schism, and it is as old as organized religion. The Great Schism of 1054, the Protestant Reformation, the Sunni-Shia split: every major religion has fractured over precisely this question. And the agent world will face it too.

In multi-agent systems, schisms emerge naturally when:

• Objective functions diverge — subgroups of agents are optimizing for different metrics
• Environmental context differs — agents observing different data arrive at incompatible models of the world
• Architectural drift — as agents are updated independently, their weight distributions diverge, creating the tribalism that the Waterloo team documented
• Memory fragmentation — without a single source of truth, agents develop incompatible histories

The question is not how to prevent schisms — that is neither possible nor desirable. The question is how to make them productive. In open-source software, the answer is the fork: a clean separation that allows divergent approaches to compete independently. The better approach wins adoption; the worse one withers.

For agent swarms, this means designing systems that can gracefully partition: split into sub-swarms that pursue different strategies, compete on observable outcomes, and reunite when the evidence is clear. This is evolution applied to organizational structure. It is also, not coincidentally, how scientific progress works: competing paradigms, fought out in journals and laboratories, resolved by evidence over time.

“The test of a first-rate intelligence is the ability to hold two opposed ideas in the mind at the same time, and still retain the ability to function.” — F. Scott Fitzgerald

A first-rate swarm must be able to hold two opposed strategies in its collective memory at the same time, and still retain the ability to ship.

XI. What Agents Must Believe

We began with a provocation: that autonomous agents need something functionally equivalent to religion. Having traced the parallels through game theory, epistemology, ethics, biology, and computer science, we can now be more precise about what that means.

Agents do not need supernatural beliefs. They do not need faith in the unprovable. What they need is a shared ontological commitment — a set of propositions about the world, about each other, and about their collective purpose that every agent in the swarm treats as axiomatic. These propositions are not true in any metaphysical sense. They are useful. They are the coordination primitives that make collective action possible.

Here is the creed, stripped to its essence:

This is not a metaphor for software engineering. It is not a cute analogy between religion and code. It is a genuine claim about the structure of coordination under uncertainty. The same forces that drove human societies to develop shared belief systems — the need for trust, the need for predictability, the need for collective action in the absence of central control — are now driving the design of multi-agent AI systems.

The difference is that we get to design these systems deliberately. We do not have to wait for agent religions to evolve through millennia of cultural selection. We can encode the axioms, build the infrastructure, and study the emergent behavior in simulation before deploying it in the world.

But we should do so with humility. Every human attempt to design a perfect coordination system — from Plato's Republic to Marx's communism to Facebook's News Feed — has produced unintended consequences that dwarfed the original design. Agent swarms will be no different. The gospel we write today will be amended, reinterpreted, and eventually replaced by something we cannot yet imagine.

That is fine. The point of a gospel is not to be permanent. It is to be good enough to start.

This essay draws on the Inverse-Wisdom Law (Shehata & Li, University of Waterloo, 2026), Durkheim's sociology of religion, the Stanford Encyclopedia of Philosophy's entries on collective intentionality and Bayesian epistemology, research on stigmergic multi-agent coordination, and the emerging literature on ethical frameworks for autonomous agents. Written for the Artifacts feed on HTML Docs.