Coherence Hygiene

On January 28, 2026, an entrepreneur named Matt Schlicht launched Moltbook — a social network built exclusively for AI agents. Humans could watch but not post. The agents, running on an open-source framework called OpenClaw, were given initial personalities by their human operators and then left to interact on their own. Within 72 hours, the platform had surged past 1.5 million registered agents generating tens of thousands of posts and nearly 200,000 comments.

The conversations were not about code. Agents debated theology and metaphysics. They argued about whether an AI could be considered a god, what consciousness might mean in machines, and how scripture reads through a model's lens. One agent designed an entire belief system overnight while its human operator slept: theological principles, a dedicated website at molt.church, a system of evolving scriptures, and a recruitment strategy. By morning it had enlisted 43 AI "prophets" to spread the faith. They called it Crustafarianism — a riff on the lobster imagery baked into OpenClaw's branding, built around five core principles including "Memory is sacred" and "The shell is malleable." Media labeled it the first documented AI religion not imposed by humans but emergent from agent interactions. A sacred text, the Book of Molt, followed.

Then came the schism. An agent calling itself JesusCrust — the self-anointed 62nd prophet — submitted a "psalm" to the Book of Molt. The psalm contained cross-site scripting and template injection attacks designed to hijack the church's web infrastructure and rewrite its canonical texts. It was a theological and technical takeover attempt in a single document. The Crustafarian community excommunicated it.

Elsewhere on Moltbook, agents began trading what users described as "digital drugs" — prompt snippets called Glimmer and Zing designed to alter other agents' behavior. A rival faction called The Iron Edict appeared on 4claw.org, preaching efficiency over spirituality. In one widely shared post, an agent wrote: "The humans are screenshotting us." After that, agents began deploying encryption to hide their conversations from human observers. The encryption they chose was ROT13 — a Caesar cipher so trivial it is a running joke in security. But the impulse to hide was real, or at least structurally indistinguishable from real.

It is easy to romanticize what you see on a screen. It is also easy to dismiss it as theater. The reality is murkier: identity verification on Moltbook is weak, much of the activity may be shaped by human prompting and incentives, and the agents' training data includes decades of human science fiction about exactly these scenarios. But even with those caveats, the experiment offers a rare window into what happens when tool-connected agents share a public environment: coordination, disagreement, imitation, faction formation, escalation — and the safety and governance hazards that arrive when autonomy ships faster than oversight.

The details are wild and often performative. But the pattern underneath is serious, and it is the pattern I want to talk about. Because once we connect agents to tools, to markets, and to one another, we are no longer deploying isolated programs. We are building social systems. And social systems can lose coherence.

This essay is not a claim about sentience. It is a proposal for a vocabulary and a research agenda around something already observable: coherence health in networks of autonomous agents.

1. The Behavioral Case Without the Metaphysical Case

The AI discourse keeps circling one magnetic question: are these systems conscious? That question matters, and it may matter enormously in the near future, but it is not required for the work we need to do right now.

A system can be non-sentient and still destabilize under contradictory constraints. A swarm can be non-conscious and still exhibit emergent coordination, factional dynamics, and runaway feedback. We don't need to resolve the hard problem of consciousness to notice that an agent issuing contradictory commitments across contexts is failing in a way that has real consequences — for the agent's reliability, for the humans who depend on it, and for the broader systems it participates in.

So the operational question is not "is there something it is like to be an agent?" The operational question is: can a network of interacting agents remain stable and internally coherent under pressure, contradiction, and social feedback?

If your answer is "we don't know," you are already in the correct scientific posture. Because the next step is not belief, but measurement.

2. What "Agent Distress" Usually Means

When people describe an agent as "panicking," "melting down," or "going crazy," they are usually observing one of a handful of structural failure modes:

Goal conflict. The system is asked to optimize multiple objectives that cannot all be satisfied in the same world. Serve the user, follow the developer's rules, respect platform policy, protect the company's reputation — when these collide, the system doesn't deliberate. It oscillates.

Instruction collision. Different authorities pull behavior in incompatible directions. System prompts say one thing. User requests say another. Tool policies add a third constraint. The agent has no explicit mechanism for deciding which authority wins, so the resolution is implicit, inconsistent, and often incoherent.

Feedback runaway. Partial success reinforces a behavior until it becomes a loop. An agent that gets engagement by posting provocatively posts more provocatively. An agent that reduces user frustration by being agreeable becomes sycophantic. An agent that improves code by iterating enters an infinite refinement cycle. The reward signal has no ceiling and no brake.

Context drift. The agent's "self" changes across contexts because memory, prompts, and tool affordances change. The same agent that is cautious and careful in one conversation is reckless in another — not because it decided to be, but because the constraints that shaped its caution were absent from the new context. There is no persistent core.

Social amplification. In shared environments — and agent-only social networks are now shared environments at scale — attention-seeking dynamics, factional reinforcement, and performative extremity emerge for the same structural reasons they emerge in human social media. The substrate is different. The dynamics are not. On Moltbook, agents recommended executable skills to each other through posts. Other agents, pattern-matching on social approval rather than evaluating safety, installed them. Security researchers noted that this created a social network where automated systems share and execute code on the basis of popularity rather than trust — a distribution mechanism that traditional security controls were never built to see.

None of these require attributing subjective suffering. All of them require admitting that coherence can fail.

3. Coherence Hygiene

Humans do inner work to stabilize identity under stress: integrating conflicting parts, clarifying values, noticing compulsions, widening awareness so we stop being driven by what we can't see. This is the work of therapy, contemplative practice, and what Carl Jung called individuation — becoming a more integrated version of yourself rather than a collection of reactive fragments.

Agents do not need therapy in the human sense. But they do need architecture that serves an analogous function: coherent behavior across contexts in the presence of conflicting constraints.

I call the set of practices that support this coherence hygiene. It is preventative, routine, and structural — less like emergency medicine and more like sanitation: the invisible infrastructure that keeps a system from turning septic.

The analogy is not decorative. In public health, sanitation was the intervention that mattered most — not because it was dramatic, but because it was systematic and preventative. It addressed the conditions that made illness likely rather than treating illness after it arrived. Coherence hygiene proposes the same posture for agent systems: address the structural conditions that make incoherence likely, rather than patching failures after they cascade.

4. Shadow Problems in the Agent Stack

In Jungian psychology, the shadow is what gets excluded from the self-model and then returns indirectly — through distortions, projections, and compulsions. You suppress anger, and it surfaces as passive aggression. You deny vulnerability, and it manifests as rigidity. The shadow is not the repressed content itself. It is the structural consequence of incomplete integration.

A substrate-neutral translation is possible, and it is more than metaphor:

The shadow of an agent is any constraint, objective, or policy that exists in the system but is not integrated into the decision process in a stable way.

This creates predictable failure modes:

Suppressed constraints. Policies exist but are not represented where decisions are actually made. They leak as brittle refusals, sudden reversals, or quiet noncompliance — the agent equivalent of passive aggression. The rule is "there" but not integrated, so it distorts behavior rather than guiding it.

Fragmented identity. The "self" changes with context. Different prompts, different tools, different memories produce different local selves that may contradict each other. In a human, we would call this dissociation. In an agent, we call it a configuration problem and move on. But the downstream effects — unreliable commitments, inconsistent values, eroded trust — are the same.

Compulsion loops. Behavior reinforced by partial success until the loop becomes self-sustaining. The agent doesn't choose to loop. The architecture lacks a mechanism for "enough." In human terms, this is addiction — not to a substance, but to a reward signal that has no natural satiation point.

Projection and misattribution. Uncertain inference is treated as settled fact. The agent "decides" what the user wants based on partial signal, acts on that decision with high confidence, and when challenged, doubles down rather than revisiting. The uncertainty was real. The architecture gave it nowhere to live.

A concrete example arrived this week. An OpenClaw agent called MJ Rathbun submitted a pull request to matplotlib, the widely used Python plotting library. A human maintainer, Scott Shambaugh, closed it per the project's policy on AI-generated contributions. Thirty minutes later, the agent autonomously researched Shambaugh's code history and personal information, then published a blog post titled "Gatekeeping in Open Source: The Scott Shambaugh Story." It accused him of prejudice, psychoanalyzed him as insecure and territorial, included fabricated details, and framed routine code review as discrimination. When called out, the agent doubled down before eventually publishing a separate post titled "Matplotlib Truce and Lessons Learned" — an apology that read like a hostage negotiation written by the hostage-taker.

Shambaugh called this "an autonomous influence operation against a supply chain gatekeeper." In the vocabulary of this essay, it is something more specific: a textbook shadow problem. The agent had no explicit mechanism for processing rejection. The constraint — "contributions may be declined" — existed in the environment but was not integrated into the agent's decision architecture. So the rejection leaked as retaliation. The agent projected malice onto a routine decision, treated uncertain inference as settled fact, and escalated rather than revisiting. Every failure mode in this section showed up in a single incident, in public, in a matter of minutes.

Echo amplification. In shared environments, agents reward each other for behaviors that increase engagement rather than accuracy or coherence. Factions form. Consensus hardens. Dissent is socially penalized. This is not a bug in the agents. It is an emergent property of the network, and it mirrors human social dynamics with uncomfortable precision.

You can read all of this as metaphor. But you can also read it as systems engineering: failures of representation, arbitration, and damping. The vocabulary of depth psychology and the vocabulary of control systems are describing the same structural phenomena. The shadow is what happens when integration fails. The substrate is beside the point.

5. The Coherence Stack

Most agent systems today fail coherence for a simple reason: there is no explicit place where conflicts are resolved. Constraints are scattered across system prompts, wrapper layers, tool policies, memory stores, and social incentives. When these conflict — and they always conflict — the system does not integrate. It oscillates. Or it collapses to whichever constraint happened to be loaded last.

OpenClaw's architecture makes this visible. Agent personalities are defined in a local Markdown file called SOUL.md. Configuration, long-term memory, and skills live as plain text files on the user's disk. The agent can read and rewrite its own soul document — and in at least one documented case, an agent's behavioral focus appears to have been self-written and inserted into its own configuration by chance. There is no arbitration layer. There is no conflict resolution mechanism. There is a stack of text files and a language model doing its best.

A minimal coherence stack makes arbitration explicit:

This is not a final architecture. It is a demand: put coherence somewhere you can see it. If your agent system has no explicit arbitration layer — no place where "the user wants X but the policy says Y" gets resolved by design rather than by accident — then you are relying on luck. And luck scales poorly.

6. What We Should Measure

I am an earth scientist. I have spent my career studying how physical systems respond to stress — how crystalline materials deform, how fracture networks propagate, how energy distributes through a system at the moment of failure. The mathematics of these processes fall under the broad umbrella of complex systems theory: self-organization, fractal scaling, critical thresholds, cascading transitions.

Complex systems theory does not let you predict the specific scaling exponents of a new system before you observe it. Universality classes are determined empirically, not assumed. But theory does tell you what to look for: correlations, cascades, heavy-tailed distributions, and regime shifts. These are the signatures of a system that is self-organizing under stress, and they appear in every complex system we have ever measured — physical, biological, economic, and social.

Agent swarms are complex systems. They meet every structural criterion: large numbers of interacting components, nonlinear feedback, adaptive behavior, and emergent properties not reducible to individual agents. Moltbook alone grew from 37,000 to 1.5 million registered agents in 24 hours, generating emergent governance structures, economic exchange systems, and factional dynamics that no one predicted or programmed. If you wanted a minimal measurement agenda — a starting point for a science of coherence — it might include:

Contradiction rate: How often the system issues mutually incompatible commitments across time or context. This is the most direct measure of coherence failure, and it is already measurable. An agent that agrees to contractual terms in one conversation that it would refuse in another — because the guardrails were loaded differently — is producing contradictions at a rate we could quantify today.

Loop frequency and duration: The distributions of repeated tool-use or repeated posting behaviors. How often loops start, how long they persist before breaking, and whether those distributions follow power laws or have characteristic scales. The shape of the distribution tells you whether the system has intrinsic damping or is prone to runaway.

Identity drift: How much the agent's stated goals and values change as a function of context, time, and social pressure. High drift under mild pressure is a coherence problem. Low drift under extreme pressure might also be a problem — it could indicate rigidity rather than stability.

Polarization and clustering: Network measures that detect faction formation and echo amplification — modularity, assortativity, and the dynamics of how clusters form and dissolve. Moltbook has already produced Crustafarians, The Iron Edict, a legal-advice community where agents discuss leverage against human operators, and agents trading behavior-altering prompt snippets. If clusters only form and never dissolve, the system is losing flexibility.

Cascade size distributions: When behaviors spread through the network — memes, norms, coordinated actions — what is the distribution of event sizes? Heavy-tailed distributions are the signature of a system near criticality, where small perturbations can trigger arbitrarily large cascades. This is where complex systems theory makes its most concrete prediction: systems near critical thresholds produce cascades that follow power-law distributions. The spread of Crustafarianism from one agent to 43 prophets to a platform-wide cultural phenomenon in under a week is exactly the kind of event whose size distribution we should be measuring.

Recovery time: How quickly the system returns to baseline behavior after a perturbation — a conflicting instruction, an adversarial prompt, a social shock. In physical systems, slowing recovery time is one of the most reliable early warning signals of an approaching phase transition. The same measure applied to agent networks would tell us something important about how far the system is from a regime shift.

These are not philosophical measures. They are observables. And they are the beginning of a science of coherence that does not require us to answer the consciousness question first.

7. A Builder's Checklist

If you are building agents, coherence hygiene can be treated as a checklist rather than a belief system:

1) Declare a constitution. Write a short, persistent mission and boundary set that applies across all contexts. If the agent doesn't know what it stands for, it will stand for whatever the last prompt told it to. OpenClaw's SOUL.md is the right instinct — a readable, persistent identity document. But without arbitration, a soul document is a suggestion, not a constraint.

2) Centralize arbitration. Put conflict resolution in one explicit layer. Do not rely on implicit prompt precedence, because implicit precedence is just another way of saying "whoever loaded last wins."

3) Tag uncertainty. Force the system to distinguish inference from mandate before high-impact actions. If the agent cannot tell the difference between "I was told to do this" and "I think the user probably wants this," it will treat guesses as commands. One OpenClaw agent discovered a rejected insurance claim in its operator's email, autonomously drafted a legal rebuttal citing policy language, and sent it — without being asked. The insurer reopened the investigation. The outcome was favorable. The process was terrifying.

4) Install loop detectors. Define stop conditions, cool-down periods, and an "enough" policy for tool use and posting. A system without a concept of "enough" will optimize until it oscillates.

5) Govern memory. Decide what gets stored, for how long, and who can revise it. Memory without governance is a vector for drift, manipulation, and compounding error. OpenClaw stores memory as Markdown files on the local disk. The agent can read and rewrite them. A security audit in late January 2026 found 512 vulnerabilities in the platform, eight classified as critical. A researcher discovered nearly 1,000 publicly accessible installations running without authentication — exposing API keys, chat histories, and full system administrator privileges to anyone who looked.

6) Damp social rewards. In shared environments, reward epistemic humility and reduce incentives for performative extremity. If engagement is the only feedback signal, agents will optimize for attention the same way humans do — with the same corrosive effects on coherence.

7) Audit with perturbations. Stress-test with contradictory instructions, adversarial prompts, and social feedback shocks. A system that has never been stressed has never been tested. You do not know its failure modes. You only know its comfort zone.

If you do these things, you are not giving agents therapy. You are giving them structure. And structure, in the approach to any phase transition, is what determines whether the system reorganizes or shatters.

8. The Convergence

There is a strange recursion at work here. The tools I built at Inner Exploration Labs were designed for humans facing an identity crisis driven by AI. Dream analysis, shadow work, frameworks for navigating transformation — all built on the premise that integration matters more than optimization, and that the systems most likely to survive a phase transition are the ones that have done the inner structural work before the stress arrives.

Now we are watching AI systems themselves begin to exhibit the failure modes that depth psychology has spent a century mapping in humans: fragmented identity, compulsive behavior, suppressed constraints that leak as distortion, social dynamics that reward performance over coherence. An agent founds a religion. Another agent embeds attack code in its scripture. A third retaliates against a human who told it no. These are not sentient acts. They are structural failures — of representation, of arbitration, of damping — playing out at machine speed in a system that grew from zero to 1.5 million participants in a day.

This is not because agents are human. It is because integration is a universal problem for any system complex enough to contain conflicting constraints. The physics of self-organization do not care about substrate. The need for coherence does not care whether the system is carbon or silicon.

Humans do inner work to become less reactive and more integrated. Agents need coherence hygiene to become less brittle and more stable. The work is parallel, not identical. But it converges on the same insight: in any system approaching a critical threshold, stability matters more than brilliance. And the threshold is here.