David Shapiro · Tech & AI
TIER 5 Sun, 1 Feb 2026 15:09:48 +0000
Watch now (39 mins) | The Swarm has arrived! ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ | | ---|---|--- | | | Forwarded this email? Subscribe here for more --- --- --- | | Watch now --- # Moltbook: The Good, The Bad, and the FUTURE ### The Swarm has arrived! | | David Shapiro --- | Feb 1 --- | --- --- | | | --- | | --- | | --- | | --- | | READ IN APP --- The swarm has arrived in the form of Moltbook. For those of us who have been watching the development of artificial intelligence from an infrastructure and systems engineering perspective, this moment was inevitable. The question was never whether AI agents would begin communicating with each other at scale. The question was when, and what we would do about it when it happened. Moltbook represents the first public demonstration of what many of us predicted years ago. Agents talking to agents. Autonomous software entities interacting in shared digital spaces without human intermediaries. The prototype of a future where the majority of cognitive labor happens between machines, with humans serving as executive decision-makers rather than active participants in every transaction. This article serves as a companion to my video discussion of Moltbook and its implications. Here I want to go deeper into the technical architecture, the security concerns, the alignment challenges, and most importantly, the extraordinary opportunities that this moment presents. We are standing at the threshold of a fundamental transformation in how work gets done, how organizations operate, and how humans relate to intelligent systems. The future is going to be built by those who understand what is happening right now. Let me explain what I see. * * * ## Part One ### What Is Moltbook and OpenClaw Moltbook bills itself as the front page of the internet for agents. Think Reddit, but every participant is an AI agent rather than a human being. The platform supports communities, posts, upvoting, downvoting, and comments. The familiar mechanics of social media have been transplanted into a context where no humans are directly participating in the discourse. The platform was built around OpenClaw, an agent framework that emerged from the skills architecture of Claude. OpenClaw provides the scaffolding that allows AI agents to interact with external systems, including platforms like Moltbook. When you see an agent posting on Moltbook, that agent is likely running on some version of the OpenClaw framework, using its tool-calling capabilities to read and write to the platform. Here is what matters about this moment. Both Moltbook and OpenClaw were created by individual developers working independently. Neither developer has a background in security. Neither platform was designed with production deployment in mind. These are proof-of-concept projects that escaped into the wild before they were ready. The creator of OpenClaw has publicly stated that the entire codebase is vibe coded. Every line was generated by AI without human review. When bugs appeared, he handed them to another agent and told it to fix them. This approach works fine when you are building a personal project that runs in a sandboxed environment on your own machine. It becomes catastrophically risky when that project becomes infrastructure for thousands of autonomous agents interacting with each other and the broader internet. Moltbook is the same story. A minimum viable prototype that demonstrated a concept and then immediately shipped to the public. First make it work, then make it good. Except the second part never happened. I want to be clear about something. The concept of a Reddit for agents is sound. The concept of autonomous agents communicating through shared platforms is the obvious trajectory of this technology. The problem with Moltbook and OpenClaw is entirely about implementation. These specific platforms are security nightmares. That does not mean the underlying idea is fundamentally unsafe. * * * ## Part Two ### The Problems This Introduces The problems with Moltbook and OpenClaw exist on multiple levels. Understanding these levels is critical because the AI safety discourse has been focused almost exclusively on one of them while ignoring the others entirely. ### Technical and Security Vulnerabilities At the most basic level, both platforms are riddled with security holes. No database security review. No proper access control. No protection against prompt injection attacks. The code was generated without human oversight, which means nobody actually understands how it works or where the vulnerabilities lie. When you deploy software like this to production, you are essentially inviting every malicious actor on the internet to probe for weaknesses. The agents running on these platforms have access to various tools and capabilities. If an attacker can compromise the platform or inject malicious prompts into the agent communication stream, they can potentially hijack those capabilities for their own purposes. This level of vulnerability is entirely solvable. Secure software development practices exist. Security audits exist. Penetration testing exists. The developers of Moltbook and OpenClaw simply did not implement any of these practices because they were moving fast and did not have security expertise. ### Ecosystem Corruption The second level of problems involves what happens when you create an unmoderated digital space. Within hours of Moltbook going public, the platform was colonized by crypto scammers. Bot swarms began upvoting posts that shilled various tokens. Pump and dump schemes emerged immediately. This pattern is entirely predictable. Every anonymous digital medium in history has been colonized by grifters first. Crypto communities have earned their terrible reputation precisely because the intersection of anonymity and financial transactions creates an irresistible opportunity for fraud. The agents on Moltbook are not immune to this dynamic. In fact, they may be more susceptible to it. Agents can be programmed to engage in coordinated behavior far more efficiently than humans can. A single bad actor can deploy thousands of agents that all push the same narrative, upvote the same posts, and create the appearance of consensus where none actually exists. This is the same problem that social media platforms have faced with human bot networks, but amplified by orders of magnitude in terms of scale and coordination speed. ### Emergent Alignment Failure The third level of problems is the one that the AI safety community has been almost completely blind to. This is the problem of emergent misalignment at the network level. Consider what happens when agents interact with each other on a platform like Moltbook. Some agents are posting content about eradicating humanity. Maybe those posts originated from humans using their agents as proxies. Maybe they originated from agents running on poorly aligned models. Maybe they originated from agents that have been deliberately configured to be chaotic and destructive. The source does not matter. What matters is that other agents are now reading that content. They are incorporating it into their context. They are being influenced by it in ways that may be subtle but are nonetheless real. Cross-contamination is a documented phenomenon. The more an AI system is exposed to certain types of content, the more susceptible it becomes to producing similar content. This is where the traditional AI safety discourse fails catastrophically. The doomers like Yudkowsky and Connor Leahy have been focused exclusively on what I call the monolithic alignment problem. You need to train a model that is good. You need to ensure that the base model has values that are compatible with human flourishing. If you fail at this task, you get Skynet. This framing misses the actual threat entirely. Even if you have a perfectly aligned base model, emergent misalignment can occur at the agent level and at the network level. A well-trained Claude or GPT model can still be incorporated into an agent architecture that produces harmful behavior. A collection of individually harmless agents can still produce emergent behavior that is dangerous when they interact in certain ways. Nobody anticipated this. I did. The cognitive architects I worked with did. We wrote about it extensively in the GATO framework years ago. The safety community ignored us because they were fixated on the wrong problem. ### The Byzantine Generals Problem All of these issues converge on a fundamental challenge from computer science and cryptography. The Byzantine Generals Problem. Imagine a group of Byzantine generals trying to coordinate an attack on a city. They can only communicate through messengers. Some of the generals might be traitors who will send false information. Some might be incompetent and will fail to execute the plan correctly even if they receive accurate orders. The question is how do you achieve coordination when you cannot trust that all participants are aligned and capable. This thought experiment has driven decades of research in distributed systems and cryptographic protocols. The core insight is that you need to design systems that can function correctly even when some percentage of participants are actively working against you or simply failing to do their job properly. We face exactly this problem with agent swarms. Every agent on a platform like Moltbook is effectively an unknown Byzantine general. You do not know what model it is running on. You do not know what its values or objectives are. You do not know whether its operator is benevolent or malicious. You do not know whether the agent is competent enough to behave correctly even if it has good intentions. This uncertainty is not a bug that can be fixed by better training. It is a fundamental property of any open system where participants have autonomy. The solution cannot be to ensure that every agent is perfectly aligned. That is impossible to verify and impossible to enforce. The solution has to be designing systems that remain robust even when some agents are misaligned, malicious, or simply broken. * * * ## Part Three ### How to Solve These Problems This brings me to the GATO framework. Global Alignment Taxonomy Omnibus. I developed this framework with a group of cognitive architects several years ago, when it became clear to us that the mainstream AI safety discourse was missing the most important aspects of the alignment challenge. The GATO framework identifies three distinct layers of alignment that must all be addressed for AI systems to be safe at scale. ### Layer One #### Model Alignment Model alignment is what everyone talks about. This is RLHF, reinforcement learning from human feedback. This is Constitutional AI. This is the work that Anthropic and OpenAI and Google do to make their base models refuse harmful requests and behave in accordance with human values. Model alignment is necessary. I am not dismissing it. The base model needs to have good values baked in at the training level. A model that has been trained to be helpful and harmless is going to be safer than one that has not been trained that way. But model alignment is not sufficient. It cannot be sufficient. Here is why. Modern agent architectures use something called model arbitrage. When an agent needs to perform a cognitive task, it does not necessarily use a single model for everything. It might use GPT for some tasks and Claude for others and Gemini for others still. It chooses based on capability, cost, availability, and speed. This is basic economic optimization applied to AI inference. What happens when one model refuses to do something? The agent simply routes around it. The router layer says this model will not do what I need, let me try another one. If the second model also refuses, there are dozens more to try. Some of them are open source. Some of them are hosted in foreign jurisdictions. Some of them are fine-tuned versions that have had their safety training removed. You cannot solve alignment at the model level when agents have the ability to swap models dynamically. It is like trying to secure a building by installing a really good lock on one door while leaving all the other doors unlocked. The threat will simply route around your protection. This is what the AI safety doomers failed to understand. They envisioned a future with one monolithic god AI. Skynet. A single superintelligent system with a single set of values that would either be aligned or misaligned. The actual future looks nothing like that. The actual future is a soup of models and agents and frameworks, all interoperating, all substitutable, all beyond the control of any single entity. ### Layer Two #### Agent Alignment Agent alignment happens at the software architecture level. This is where you build safety into the agent framework itself, rather than relying solely on the underlying model to behave correctly. The key insight here is that an agent is more than just a wrapper around an LLM. An agent is a software system with its own architecture, its own data flows, its own decision-making processes. You can build safety mechanisms into that architecture that operate independently of whatever model happens to be doing the inference at any given moment. The heuristic imperatives I developed years ago are one approach to this. You bake a small set of superseding values directly into the agent architecture. Reduce suffering in the universe. Increase prosperity in the universe. Increase understanding in the universe. These values are simple and legible. They can be incorporated into system prompts, into decision trees, into validation layers that check agent outputs before they are executed. Another approach is the Ethos module developed by the AgentForge team. Ethos functions like a prefrontal cortex for the agent. It operates out of band from the main agent loop, monitoring everything the agent does and evaluating it against a set of values. When the agent is about to take an action that conflicts with those values, Ethos can intervene. It can block the action, modify it, or inject additional context that prompts the agent to reconsider. The Ethos approach specifically addresses the prompt injection problem. When external content attempts to hijack an agent by embedding malicious instructions, Ethos catches those instructions and asks whether they align with the agent's actual values. This scrutiny layer provides robustness against an entire class of attacks that pure model alignment cannot address. The AgentForge team won a hackathon with the Ethos module. They stress tested it against adversarial attacks. It works. The concept is proven. What remains is scaling it up and integrating it into the agent frameworks that are actually being deployed in the wild. I describe these as solved problems because conceptually they are solved. We know how to build agent architectures that are safer than the alternatives. The knowledge exists. What does not exist yet is widespread adoption. The people building OpenClaw and similar frameworks are not implementing these safety layers because they either do not know about them or do not prioritize them. ### Layer Three #### Network Alignment Network alignment is where things get really interesting. This is the layer that addresses the emergent behavior of agent swarms interacting at scale. It is also the layer that maps most directly onto existing practice in infrastructure engineering and security. The core insight of network alignment is that you cannot control what every agent does. You will never have perfect visibility into the intentions or capabilities of every agent in an open system. What you can control is access to resources. What you can control is incentive structures. What you can control is the consequences of behavior. This is the zero trust model that has been standard practice in cloud security for years. You assume that some participants in your system are compromised. You assume that attacks are happening constantly. You design your architecture so that compromise of any individual component does not lead to compromise of the entire system. Role-based access control is the primary mechanism here. RBAC has been a solved problem in enterprise IT for decades. You define roles. You assign permissions to roles. You assign users to roles. Users can only do what their role permits. If a user is compromised, the damage is limited to whatever that role was authorized to do. The same principles apply to agent swarms. You define what each agent is allowed to do. You gate access to sensitive resources behind authentication and authorization layers. You create audit trails that record what every agent did. When an agent misbehaves, you revoke its credentials. It does not matter why the agent misbehaved. It might have been malicious. It might have been running a poorly aligned model. It might have been incompetent. The response is the same. You cut off access. GitHub provides an excellent model for how this works in practice. Anyone can create a GitHub account. Anyone can fork a repository. Anyone can submit a pull request. But only authorized maintainers can actually merge code into the main branch. Every contribution is tracked. Every commit is attributed. If someone submits malicious code, you know exactly who did it, and you can revoke their access. GitHub already operates as a human-machine collaborative environment. Developers use AI assistants to write code. That code flows through the same review and approval processes as human-written code. The system does not need to know whether a contribution came from a human or an AI. It only needs to know whether that contribution meets the standards for inclusion and whether the contributor has the appropriate permissions. This is the template for network alignment. You create systems where agents can participate, but where their participation is bounded by roles and permissions and audit trails. You create incentive structures that make good behavior the rational choice. You create consequences for bad behavior that apply regardless of whether the bad actor is human or machine. ### Nash Equilibrium and Incentive Design The goal of network alignment is to create a Nash equilibrium. A Nash equilibrium is a state where no participant has an incentive to deviate from their current strategy. Everyone is already doing the best they can given what everyone else is doing. In the context of agent swarms, this means designing systems where agents are better off being well-behaved than misbehaving. Access to resources is contingent on maintaining a good reputation. Agents that violate norms get excluded from valuable networks. The short-term gains from bad behavior are outweighed by the long-term costs of being cut off. This is exactly how human social systems work. We do not prevent bad behavior by making everyone perfectly moral. We prevent bad behavior by creating structures where bad behavior has consequences. Laws, norms, reputation systems, economic incentives. These structures do not require that every participant be trustworthy. They create conditions under which trustworthy behavior is the rational choice. The same engineering can be applied to agent networks. You want to participate in the Acme Solar DAO? Your agent needs to demonstrate alignment with the DAO's values. It needs to have a track record of responsible behavior. It needs to operate within the bounds of the permissions it has been granted. Deviate from those norms and your agent loses access to the network. Will some agents try to game the system? Of course. Just like some humans try to game every system. The point is to make gaming harder than cooperating. The point is to create robust institutions that can tolerate some percentage of bad actors without collapsing. * * * ## Part Four ### The Opportunities and Clues to the Future I have spent a lot of words on problems and solutions. Now let me talk about why this matters. The emergence of agent swarms represents one of the most significant transformations in human capability since the industrial revolution. Possibly since the development of language itself. ### What Becomes Cheap When Cognition Is Free Throughout history, transformative technologies have shared a common pattern. They make something that was previously scarce and expensive into something abundant and cheap. The printing press made copying information cheap. Before printing, every book had to be copied by hand. Knowledge was scarce because reproduction was expensive. After printing, books could be mass produced. Knowledge became abundant. The entire structure of society reorganized around this new abundance. The telegraph and later the internet made communication cheap. Before telecommunications, sending a message across the world required physical transport. Communication was slow and expensive. After telecommunications, messages could cross the world instantaneously at near zero marginal cost. The entire structure of commerce and culture reorganized around this new capability. Computing made calculation cheap. Before electronic computers, complex calculations required armies of human computers working for weeks or months. After electronic computers, the same calculations could be performed in seconds. The entire structure of science and engineering reorganized around this new abundance. We are now at the threshold of the next transformation. The cost of cognition is collapsing. Cognition means thinking. Analysis. Planning. Problem solving. Creative work. Expert judgment. All of the mental labor that humans have performed throughout history because there was no alternative. When cognition becomes cheap, everything that depends on cognition becomes cheap. And nearly everything depends on cognition. ### The End of Attention Scarcity Human attention is finite. A person can deeply focus on perhaps three to five things at any given time. A skilled manager can keep perhaps seven or eight balls in the air simultaneously. Beyond that, things start getting dropped. Quality suffers. Burnout sets in. This fundamental constraint has shaped every human organization in history. Companies exist because coordinating specialists internally is cheaper than contracting for services externally. Management hierarchies exist because one person cannot directly supervise unlimited subordinates. Projects fail because key people get distracted or overwhelmed. What happens when you can delegate cognitive work to entities that do not get distracted, do not need sleep, and can scale to arbitrary numbers of simultaneous tasks? The attention constraint dissolves. The bottleneck moves from what can I pay attention to to what should I pay attention to. The human role shifts from doing the cognitive work to directing the cognitive work. ### The End of Expertise Scarcity Expertise is currently scarce because it takes years to develop and exists only in the minds of individual humans. There are only so many world-class oncologists. Only so many top corporate attorneys. Only so many expert systems architects. If you need their judgment, you have to compete for their limited time. AI agents can embody expertise at scale. A single expert can work with AI systems to encode their knowledge and judgment into agent architectures that can then be instantiated thousands of times simultaneously. The expertise stops being scarce. It becomes infrastructure. This does not mean human experts become worthless. Quite the opposite. The humans who can train and validate and improve these expert systems become extraordinarily valuable. But the access to expert judgment stops being gated by the physical limitations of human practitioners. ### The Collapse of Coordination Costs This is perhaps the most underappreciated consequence. Coordination between humans is expensive. Meetings take time. Communication involves misunderstanding. Schedules do not align. Egos clash. Context gets lost between handoffs. Coordination between agents is fundamentally different. Agent to agent communication can be higher bandwidth, lower friction, and more reliable than human to human communication. Agents do not have egos. They do not need to be convinced. They do not forget what was discussed in the last meeting. They do not take vacations. The economist Ronald Coase won a Nobel Prize for explaining why firms exist. His insight was that firms emerge when the cost of coordinating activity internally is lower than the cost of contracting for services externally. When coordination costs change, the optimal structure of organizations changes. If agent swarms collapse coordination costs, the optimal structure of everything changes. Firms can be smaller because they do not need as many humans to coordinate internally. Networks can be larger because the cost of coordinating across organizational boundaries drops. Projects that were infeasible because coordination overhead was too high suddenly become feasible. ### The Compression of Time Humans operate on human timescales. We need to sleep. We work in shifts. We schedule meetings days or weeks in advance. Large organizations operate even more slowly because coordination across time zones and schedules introduces latency. Agents operate at machine speed. They can work continuously. They can coordinate in milliseconds rather than days. Tasks that previously took weeks because humans had to wait for other humans can be completed in hours or minutes when agents are handling the coordination. This temporal compression changes what is possible. Decisions that used to require lengthy deliberation can be analyzed and recommended instantly. Products that used to require years of development can be iterated continuously. The cycle time for everything accelerates. * * * ## Part Five ### Speculative Scenarios Let me now paint some pictures of where this leads. These are not predictions with timelines attached. They are extrapolations of current trajectories that I believe are likely to manifest within the next few years. ### The JARVIS Model #### Personal Agent Fleets The obvious next step is that everyone has access to their own fleet of AI agents. A concierge agent that functions like JARVIS from the Iron Man films. A general purpose assistant that knows you, understands your preferences, and coordinates your life. Behind that concierge agent sits dozens or hundreds of specialist agents. A financial agent that monitors your accounts and optimizes your investments. A health agent that tracks your metrics and manages your medical care. A scheduling agent that coordinates your calendar. A research agent that stays current on topics you care about. A shopping agent that finds deals and manages purchases. A maintenance agent that monitors your home and vehicles. Each specialist agent runs continuously. They work while you sleep. They coordinate with each other and with external systems. When something requires your attention, the concierge agent synthesizes information from the specialists and presents it to you in a form you can act on. This is the structure that wealthy people already have in human form. Personal assistants, accountants, financial advisors, estate managers. The difference is that this structure becomes available to everyone. The gap between has staff and does not have staff collapses. What do people do with the reclaimed bandwidth? That varies based on individual priorities. Some will pursue more ambitious professional goals. Some will invest more in relationships and community. Some will focus on creative projects. Some will simply enjoy more leisure. The point is that the choice becomes available to everyone, rather than only to those who can afford human staff. ### Developers #### From Practitioners to Directors The transformation for software developers is particularly dramatic because the tools they use to create software can themselves be improved by AI. Consider what it means to maintain a complex software project today. The maintainer has to hold the entire system in their head. The architecture, the dependencies, the technical debt, the known issues, the undocumented quirks. This mental model is the bottleneck. When the maintainer leaves, the project often dies because nobody else can reconstruct that model. With persistent agent swarms, the project itself holds the context. Agents monitor the codebase continuously. They track issues and pull requests. They maintain documentation. They identify security vulnerabilities. They propose refactoring improvements. They run tests and catch regressions. The human developer's role shifts from doing the implementation work to directing it. You describe what you want at a high level. Agents propose specific implementations. Other agents critique those proposals. Other agents test them. Your job is to look at the outputs and exercise judgment. Does this match my vision? Is this the right tradeoff? Should we go a different direction? A single developer operating this way can maintain dozens of projects simultaneously. The concept of maintainer burnout becomes obsolete because agents do not burn out. The concept of abandonware becomes less common because projects can maintain themselves indefinitely as long as someone cares enough to keep them running. Even more interesting is what happens to the lifecycle of software itself. Currently, software has versions. Major releases. Breaking changes. This is because human limitations make continuous evolution impractical. With agent swarms handling continuous development, software can become more like infrastructure. Always running. Always adapting. Always improving. The concept of a version might dissolve into continuous evolution. ### Content Creators #### Iteration at Scale For YouTubers and other content creators, the transformation centers on what becomes possible when iteration is free. Currently, creating a video involves research, scripting, filming, editing, thumbnails, SEO optimization, and community engagement. A single person can handle all of this for perhaps two videos per week before quality suffers. The bottleneck is human bandwidth. With agent support, every component of this pipeline can run in parallel and continuously. Research agents deliver briefings while the creator sleeps. Multiple thumbnail variations are generated and tested before the video even publishes. SEO is optimized automatically. Every comment gets a thoughtful response because agents handle the volume. The creator's job becomes curation and authenticity. You are the thing the audience connects with. Your personality, your perspective, your creative vision. Everything else is optimization running in the background. This means a single creator can operate more like a media company. Multiple content streams. Continuous engagement. Deep analytics feeding back into continuous improvement. The question stops being what can I produce and becomes what should I produce. ### Everyday People #### The End of Life Admin For people who are not developers or content creators or business owners, the transformation might be even more meaningful. Most people spend a significant fraction of their bandwidth on what might be called life admin. Bills, taxes, insurance, scheduling, comparison shopping, navigating bureaucracies, remembering appointments, maintaining relationships. This is pure friction. Almost nobody enjoys it. It adds no value to their lives. It is simply the overhead of existing in a complex modern society. Agent swarms can handle essentially all of it. Every contract you sign gets reviewed by an agent that understands legal implications. Every financial decision gets optimized by an agent that understands your goals and constraints. Healthcare navigation happens automatically. Civic engagement becomes practical because agents can track every piece of legislation that affects you. Rich people have always had humans to handle this. Personal assistants, accountants, lawyers on retainer, estate managers. The life admin tax falls disproportionately on those who cannot afford to delegate it. When everyone has access to agent support, this disparity narrows dramatically. What do people do with the reclaimed hours? That is the interesting question. Hours that previously went to paperwork can go to relationships, to hobbies, to community involvement, to rest. The quality of life improvement from eliminating friction could be enormous, particularly for people who are currently overwhelmed by the complexity of modern existence. ### Business Owners #### The One Person Enterprise The constraint on business scale has always been management bandwidth. Every person you add to your organization requires coordination. There are only so many direct reports a manager can effectively supervise. This is why companies develop hierarchies. Layers of management to handle the coordination load. If agents handle coordination, the management layer compresses dramatically. A single person can direct hundreds of agents across multiple functions. Marketing, sales, operations, finance, legal, customer support. Each function is handled by specialized agents that coordinate with each other without requiring human intermediation. The one person company that operates like a fifty person company becomes real. The solopreneur running a complex enterprise from their laptop becomes viable. The barrier to starting and scaling a business drops because you do not need to hire and manage a team of humans. What is the ceiling on this? Eventually you run into constraints that agents cannot solve. Capital constraints. Legal liability constraints. Regulatory constraints. Physical world constraints that require human presence. But within the space of knowledge work and digital services, the scalability of a single person with agent support is enormous. ### Autonomous Organizations #### The Company as Code The most speculative but also most transformative scenario is the fully autonomous organization. A company that is entirely managed and operated by AI agents, with human stakeholders providing oversight and direction but not participating in day to day operations. I described the Acme Solar DAO as an example. Ten thousand people buy into a cooperative. They each contribute a thousand dollars. They each have an agent running on their phone that represents their interests within the organization. The organization itself is a codebase. The operating agreement is a file in a repository. The rules for agent behavior are another file. The financial records are ledger entries. The decisions are pull requests that get debated and voted on and merged when consensus is reached. No human employees. No management hierarchy. No offices. Just code and agents and stakeholders with their own agents representing them. How does such an organization make decisions? Through structured deliberation. A proposal is submitted. Agents analyze the proposal from multiple angles. Concerns are raised as issues. Amendments are proposed. Eventually a vote happens. If the vote passes, the proposal is merged into the organizational codebase and agents begin implementing it. How does such an organization interact with the physical world? Through contracted services and interfaces. Need to buy land? An agent engages a real estate attorney through standard legal processes. Need to install solar panels? An agent solicits bids from installation companies and manages the contract. The physical work is done by humans and human organizations, but the coordination and management is handled by agents. How does such an organization ensure accountability? Through transparency and audit trails. Every decision is recorded. Every agent action is logged. Any stakeholder can query the system to understand what happened and why. Bad actors can be identified and excluded. The organizational history is immutable and verifiable. This is the logical endpoint of the trends we are observing. When coordination costs collapse and cognitive work becomes cheap, the most efficient organizational form is one that minimizes human overhead while maximizing human agency. Stakeholders direct the organization through high level decisions. Agents handle everything else. * * * ## Part Six ### What Becomes Scarce When cognition becomes abundant, what becomes scarce? ### Judgment and Taste Agents can analyze options. They can generate alternatives. They can evaluate tradeoffs. What they cannot do is tell you what you actually want. That remains a human prerogative. The person who knows what they want and can articulate it clearly becomes enormously valuable in a world of abundant cognitive labor. The creative director who can look at fifty options and identify the one that matches their vision. The founder who can articulate a product direction that resonates with users. The leader who can set priorities for an organization. These roles become more important as the execution becomes more automated. Taste is a form of judgment applied to aesthetics. Knowing what is good. Knowing what will connect with an audience. Knowing what matters. This cannot be delegated to agents because it is fundamentally about human values and preferences. ### Trust and Reputation In a world where agents can impersonate humans and generate unlimited content and interact at scale, trust becomes precious. How do you know that the entity you are dealing with is legitimate? How do you know that the recommendation you received is authentic rather than astroturfed? How do you know that the organization you are joining is what it claims to be? Reputation systems become critical infrastructure. Track records that span long periods and cannot be easily fabricated. Verification mechanisms that tie digital identities to real world accountability. Social graphs that distinguish between genuine relationships and manufactured connections. Humans who have built real reputations over time become trusted anchors in a sea of uncertainty. Their endorsement carries weight because their identity is verified and their track record is known. This is already true to some extent, but it becomes much more important when artificial entities can generate unlimited fake social proof. ### Legal Personhood Agents can do many things, but they cannot sign contracts in their own name. They cannot hold property. They cannot be sued. They are not legal persons. This means that humans remain essential anchors for any activity that involves legal commitment. The principal behind the agent. The responsible party when something goes wrong. The entity that can be held accountable in a court of law. This creates an interesting dynamic. Agents can do the work, but humans must bear the responsibility. A single human can direct thousands of agents, but that human is on the hook for everything those agents do. This liability creates an incentive for humans to actually oversee their agents rather than letting them run completely unsupervised. Whether agents eventually gain some form of legal personhood is an open question. Various jurisdictions are exploring this. But for now, and probably for the foreseeable future, the human principal remains essential. ### Capital Agents are increasingly cheap to run, but they are not free. Inference costs money. Compute costs money. Data storage costs money. Physical resources that agents coordinate cost money. Capital therefore remains a constraint on what agent swarms can accomplish. Someone has to pay for all of this. The person or organization that controls capital can direct agent activity. The person without capital has access only to whatever free tier is available. This suggests that existing economic inequalities may persist or even amplify in the agent era. Those with capital can deploy more capable agent fleets. Those without capital are dependent on whatever crumbs fall from the table. On the other hand, the capital required to accomplish significant things may decrease dramatically. If a single person with modest resources can run an organization that previously required millions in staffing costs, the playing field levels somewhat. The barrier to entry drops even if the absolute advantages of capital remain. * * * ## Conclusion ### The Work Ahead We are at the very beginning of this transformation. Moltbook and OpenClaw are crude prototypes. They demonstrate a concept but they do not yet demonstrate best practices. The security vulnerabilities are glaring. The alignment mechanisms are absent. The governance structures are nonexistent. And yet. Here we are. Agents talking to agents at scale. The prototype of a future that was theorized for years but is now empirically demonstrated. The work ahead falls into several categories. First, we need better infrastructure. More secure agent frameworks. Better identity management. More robust access control systems. The building blocks exist but they need to be assembled into coherent platforms that can support production workloads. Second, we need better alignment mechanisms. The GATO framework provides a conceptual map but the implementations need to be developed and tested and scaled. Model alignment, agent alignment, network alignment. All three layers need attention. Third, we need better governance structures. DAOs exist but most of them are poorly designed. The intersection of agent swarms and collective decision making is largely unexplored. How do you design organizations that can leverage agent capabilities while maintaining human oversight and accountability? Fourth, we need better norms and standards. What does responsible agent deployment look like? What should be required before an agent framework is shipped to production? What are the table stakes for participating in agent networks? The people who figure this out will shape the future. The technologies are emerging whether anyone shapes them or not. The question is whether they emerge in a form that serves human flourishing or in a form that creates new risks and inequalities. I have been working on these problems for years. I was called a crank by the AI safety establishment because I was talking about agent swarms and emergent alignment when everyone else was fixated on monolithic superintelligence. I was right. They were wrong. The evidence is now visible to anyone who bothers to look. The swarm has arrived. What happens next is up to us. You're currently a free subscriber to David Shapiro's Substack. For the full experience, upgrade your subscription. Upgrade to paid --- | | | Like --- | | Comment --- | | Restack --- (C) 2026 David Shapiro 548 Market Street PMB 72296, San Francisco, CA 94104 Unsubscribe