Toward a relational infrastructure for work in the age of artificial agents

[Version française]

Abstract

The first nine issues of the newsletter Humans in the Loop develop a central hypothesis: the core problem of digital work is no longer only access to tools, platforms, data, or artificial intelligence models. The problem is the loss of collective frameworks capable of turning interactions into relationships, conversations into memory, unclear needs into action, and individual AI use into shared learning. Social platforms optimize visibility, messaging tools accelerate circulation, collaboration suites organize productivity, and workflows model processes. Yet none of these logics fully answers the relational question: how can humans remain able to help one another, recognize one another, supervise agents, transmit what they learn, and keep track of what has been decided? This article frames BouclePro as a relational infrastructure: an architecture of small Loops where humans, AI systems, traces, skills, needs, and decisions can remain collectively governable. It also develops the emerging figure of the AI Shepherd: not a mere AI user, nor a passive validator, but a human actor able to formulate, supervise, verify, connect, transmit, and maintain collective memory.

Keywords: human-in-the-loop, AI Shepherd, BouclePro, CyberWorkers, weak ties, communities of practice, collective intelligence, human oversight, collective memory, agentic AI.

Introduction: work has left the office, but has it recovered the collective?

In 1996, CyberWorkers carried an intuition that has become obvious thirty years later: work was going to leave the office. At that time, the challenge was to make teleworkers, freelancers, and remote professionals visible. It was about building bridges between needs and skills in a world where remote work was still marginal, almost experimental.

In 2026, the question is different. Work has indeed left the office. Messaging systems, social networks, collaborative platforms, learning spaces, and personal AI assistants have transformed how people produce, learn, communicate, and make themselves visible. Yet this new power has not automatically rebuilt collective life.

We have gained tools, but often lost shared context. We have gained automation, but not always cooperation. We have gained personal AI, but may lose organized collective intelligence. This is the paradox: each person can become faster, more productive, better assisted, and still become more alone inside their own cognitive loop.

The thesis of this article is straightforward: the future of work will not depend only on the power of AI models. It will depend on our ability to create human loops that produce relationships, memory, supervision, and transmission. In other words, the challenge is not merely to keep humans in the loop of AI. It is to create loops where humans remain able to think, decide, and learn together with AI, without being absorbed by it.

1. Connection does not create relationship

The first misunderstanding of digital work lies in a confusion: connecting people does not create a relationship. Platforms can recommend profiles, measure engagement, track clicks, analyze comments, and detect professional proximity. Yet they still poorly understand the fragile moment when an interaction becomes a relationship.

Mark Granovetter’s work on the strength of weak ties remains decisive here. Important opportunities do not always come from close relationships. They often emerge from peripheral connections, from weak bridges between distinct social worlds. A study by MIT, Harvard, and Stanford using LinkedIn data confirmed this intuition at scale by showing the professional value of weak ties in the circulation of opportunities.

But weak ties are not only structural connections. They are also made of weak human signals: humor, irony, self-mockery, relational context, emotional timing, and shared implicit meanings. A sentence, a tone, a joke, or a detail can open a relationship. This is precisely what platforms still struggle to interpret. Farah Benamara’s work with CNRS on humor, emotions, and online discourse reminds us that the meaning of a text depends heavily on context. A machine may detect words while missing tone, relationship, and situation.

The issue is therefore not only algorithmic. It is anthropological. Platforms can capture attention, but they do not adequately support the birth of a bond. They analyze past behaviors, but understand present states poorly. They expose profiles, but do not always protect what begins to take shape between two people.

A work infrastructure for the AI age should therefore not simply optimize connection. It should help what emerges from chaos survive long enough to become real: a conversation, a mutual aid situation, a mission, a transmission, a trust relationship.

2. A collective cannot be measured only by activity volume

The second misunderstanding concerns the life of collectives. Digital tools often confuse activity with vitality. A group that publishes a lot seems alive. A silent group seems dead. In reality, the situation is more complex.

A collective can die silently. Members remain present, but read less. They mute notifications. They tell themselves they will come back later. Then they no longer dare to return because they have missed too many exchanges and no longer know where to resume. The psychological cost of returning increases as the thread moves forward without them.

Conversely, a collective can remain alive with few messages if people know they can come back, ask for help, offer a skill, be recognized, and contribute without having to perform socially. The right indicator is therefore not only the number of posts. It is the collective’s ability to preserve a context of belonging, trust, and usefulness.

This intuition resonates with research on groups as resources for health and resilience. The World Health Organization Commission on Social Connection has made isolation a global public health priority. Julianne Holt-Lunstad’s research on social relationships and mortality risk shows that the quality of social relationships is not just an emotional supplement. It has measurable effects on health. In a similar line, the social cure approach developed by Jolanda Jetten, Catherine Haslam, and Alexander Haslam shows that belonging to groups can protect people in periods of transition, disruption, and vulnerability.

For a person in professional recovery or transition, a collective is not an audience. It is a space where they count. A place where they can become visible differently than through social performance. A place where silence is not immediately interpreted as absence.

3. Silence is sometimes an unformulated request

In digital environments, silence is often misread. Not replying becomes not being engaged. Not posting becomes not contributing. Not reacting becomes not being present. This reading is too poor.

Silence can be cognitive fatigue. Caution. Fear of judgment. Lack of social energy. Difficulty formulating a request. It can also be a discreet form of presence: someone is reading, following, observing, but does not yet have the words, confidence, or strength to enter the exchange.

This is where conversational AI changes the situation. Many people speak more easily to a machine than to a human group. Not because the machine is better than a human, but because it costs less emotionally. It does not appear to judge, it responds immediately, reformulates, and sometimes gives the impression of being available at the right moment. This experience extends, in a new form, what has long been known as the ELIZA effect.

But this help carries a risk. Each person can become enclosed in their own cognitive loop: their AI, their prompts, their projects, their doubts, their answers. Reports about behavioral collection systems for AI training, including those relayed by Korben, remind us that our actions, corrections, and conversations can also become raw material. Jaron Lanier’s critiques of free and extractive digital culture remain highly relevant here.

The issue is not to oppose individual AI use and collective intelligence. The issue is to create a passage between them. What each person clarifies with AI should be able to become shareable, discussable, verifiable, and transmissible in a Loop.

4. The future of work may lack methods more than tools

We speak constantly about the future of work, yet often underestimate a simple question: what do we produce together? After a discussion, what remains? A thread of messages, good energy, interesting remarks, or a decision, a trace, an action, a learning outcome?

The problem of digital work is not only a lack of tools. The tools already exist: messaging, video calls, personal AI, shared documents, social platforms, dashboards, CRMs, workflows. What is often missing are collective methods for turning a conversation into a decision, a decision into action, and action into memory.

This is why a Loop cannot be reduced to an AI-augmented conversation. It must rely on methods of questioning and deliberation. The Socratic maieutic method reminds us that thought can be brought out through questioning rather than immediate answers. David Bohm’s dialogue invites participants to suspend automatic reaction and observe the collective field. Edward de Bono’s lateral thinking and Six Thinking Hats help explore multiple angles without reducing discussion to opposing opinions. F. David Peat’s gentle action invites intervention in complex systems without crushing them. Robert and Michele Root-Bernstein’s thinking tools remind us that creativity is cultivated through observation, analogy, modeling, play, transformation, and synthesis. Finally, Climate Fresk shows that a simple, facilitated, reproducible, and collective protocol can transform a complex subject into shared learning.

The future of work therefore needs more than AI. It needs frameworks that help humans question better, slow down better, explore better, decide better, and transform exchanges into action.

5. The cyberworker becomes a distributed human

The cyberworker of 1996 was a connected worker: someone able to leave the physical office, sell skills online, and work remotely. In 2026, the term becomes deeper. The cyberworker becomes a distributed human: extended by tools, agents, traces, prompts, digital memories, and conversations with artificial systems.

This transformation is not only functional. It changes presence. Eduardo Kac’s telepresence artwork Rara Avis already made it possible, in 1996, to understand virtualization not as simply putting something online, but as a displacement of viewpoint, body, action, and relation. In the same art-science horizon, Roger F. Malina’s work on CaravanserAI and the Leonardo ecosystem remind us that the future of intelligence must be considered through art, science, technology, culture, bodies, languages, and memories.

The image of the caravanserai then becomes essential. In a caravanserai, travelers do not merely pass through. They leave stories, maps, dangers, methods, news. They catch their breath. They leave differently.

A Loop can be understood in this way: a professional caravanserai in the age of AI. People do not always arrive with a clear request. They arrive with an unclear question, a skill, fatigue, doubt, a desire to help. In this vision, AI is not the center. It is the discreet steward of the place: it helps translate, summarize, connect, recall, contextualize, and bring the right people to the surface at the right time. But it does not replace human hospitality.

6. Human in the loop: guarantee or exploitation?

The term human-in-the-loop is ambiguous. It can name a guarantee: humans retain control, validate, correct, contextualize. But it can also hide exploitation: the human becomes the invisible cog who clicks, annotates, filters, trains, and validates without real recognition.

Even the most advanced AI systems still depend heavily on human feedback. Tools such as Label Studio and Argilla organize annotation, evaluation, and correction workflows. InstructLab, backed by IBM and Red Hat, proposes a community-oriented logic for contributing knowledge to open source models. European law also recognizes this requirement: the CNIL distinguishes several human control regimes, and Article 14 of the European AI Act requires human oversight for certain high-risk systems.

The real question therefore becomes: how will these humans be treated? As invisible, precarious, atomized workers? Or as recognized members of a learning community? The International Labour Organization’s work on digital labour platforms and Antonio Casilli’s investigation of click work show the risk of an AI economy fed by invisibilized workers.

By contrast, a contributive model must recognize inputs, document contributions, support skill development, and define an economic framework when a contribution becomes structured work. This is where references to open source commons, the Peer Production License, and platform cooperativism become useful: a commons can be open without becoming extractive.

This question matters particularly for Francophone Africa, the Maghreb, and the Indian Ocean region. Initiatives such as Zindi, Open AIR, ReFFAO, and RFFLabs show that another model is possible: learning by doing, peer-to-peer transmission, open innovation, collaborative commons. The goal is not to build for these territories, but with them.

7. If you are human, you are the product

For a long time, we repeated: if it is free, you are the product. With AI, this sentence is no longer sufficient. Even those who pay can become the product.

We no longer commercialize only our attention, our data, or our time. We expose our questions, hesitations, formulations, corrections, and need to be understood. Part of our humanity becomes raw material. The issue is therefore no longer only: what can AI do in our place? The real issue becomes: within what human framework do we want to use it?

In an extractive framework, each person converses alone with a machine that learns from them. In a relational framework, what each person learns with AI can be shared, discussed, verified, corrected, and transmitted. The difference is not merely technical. It is institutional.

BouclePro becomes meaningful precisely here. The aim is not to add AI to a community platform. It is to create an environment where AI helps transform individual expression into a collective resource, without reducing humans to sources of data, clicks, or corrections.

8. The feed does not remember

The closure of Meta Workplace is not just an industrial episode. It is the symptom of a deeper failure: the idea that a social feed can organize work in common. Analyses by TechCrunch, Reworked, and IT Pro converge on a key point: Meta misunderstood the complexity of collective work, especially the importance of documents, trust, security, and organizational memory.

A news feed does not make an organization cooperate. It occupies it. It stimulates, accelerates, makes visible, but forgets. It scrolls content, but does not spontaneously transform requests, offers, decisions, and learning into common memory.

The tool landscape reveals an empty space. LinkedIn optimizes visibility. Meta optimized engagement. Microsoft optimizes productivity. SAP and IBM optimize processes. But who optimizes cooperation, trust, and accumulated memory at a small scale?

A Loop must answer that void. It does not merely circulate messages. It keeps track of what has been asked, offered, learned, and decided. AI is not used to produce more content, but to transform traces into common memory. The feed forgets. The loop remembers.

9. Cogs or shepherds: the figure of the AI Shepherd

In the age of AI agents, a new human role appears. It is no longer only about knowing how to prompt. The prompt is only an entry point. One must clarify an intention, choose the appropriate tool or agent, observe the result, question the answer, request verification, reintroduce other humans when necessary, and keep track of what has been decided.

This role can be called the AI Shepherd. The term is not yet stabilized in academic or regulatory literature. The dominant notions remain human-centered AI, human oversight, guided autonomy, traceability, or human-in-the-loop. But the word Shepherding names an emerging function: guiding AI agents without becoming their servants, owners, or invisible cogs.

This function is already required by governance frameworks. The NIST AI Risk Management Framework: Generative AI Profile, the European AI Act, the UNESCO Recommendation on the Ethics of Artificial Intelligence, and Stanford HAI’s work on AI in the loop converge around one requirement: humans must remain able to supervise, understand, challenge, and govern systems.

But the AI Shepherd is not only a matter of compliance. It is also a matter of learning. Anthropic’s resources on effective agents, OpenAI’s resources on building agents, and research on practical agent autonomy show that agentic environments require new skills: orchestration, verification, memory, control, audit, and critical judgment.

Research on human judgment confirms this. The work of Eren Bilen and Justine Hervé shows that bad AI advice can reduce human performance when users do not examine it critically. Supervision cannot therefore be decorative. It must become a practical skill.

The Shepherd is the opposite of the invisible cog. They are not a hand clicking in the shadows. They are a situated actor able to formulate, verify, connect, correct, transmit, and keep the thread.

10. BouclePro as relational infrastructure

For this vision to hold, it needs a clear architecture. BouclePro distinguishes Platform, Organization, Loop, Member, and Interaction. The Organization is the boundary for governance, security, data, and billing. The Loop is the relational, collaborative, contextual, and operational group inside an Organization.

This distinction is essential. If the Loop were merely a discussion group, it would fall back into the problem of the feed. If it were a full organization, it would become too heavy. Its strength is to be an intermediate unit: small enough to preserve trust, structured enough to produce memory, open enough to host humans and AI in a shared context.

BouclePro can therefore be understood as relational infrastructure. Its function is not only to connect people, nor to add AI to a community. Its function is to make the interactions between humans, agents, traces, needs, skills, decisions, and learning collectively governable.

A good Loop has three functions. It has a relational function: it helps weak ties become useful ties. It has a cognitive function: it transforms exchanges into memory and learning. It has an ethical function: it makes contributions visible and limits invisible extraction.

Only under these conditions can AI remain in its right place: facilitator, coordinator, memory, clarification aid, synthesis tool, but not the symbolic center of the collective. The center remains human, not in the abstract sense of ethics charters, but in the concrete sense: people who ask, help, doubt, transmit, verify, learn, and regain their place in a collective.

Conclusion: keeping the thread

The first nine issues of Humans in the Loop tell a clear story. It begins with a critique of connection: connecting profiles does not create relationships. It then moves to the question of the collective: a group can die silently if context, trust, and the ability to return disappear. Then comes the question of AI: it can help individuals formulate, but it can also trap each person inside their own cognitive loop. Then comes the need for methods: without a framework, conversations produce neither deliverables nor learning. Then the image of the distributed cyberworker requires us to think about professional caravanserais: spaces of hospitality, memory, and transmission in the age of AI.

Finally, the question becomes political: will humans in the loop be invisible cogs, exploited by systems they train without recognition, or shepherds capable of guiding agents, supervising decisions, and preserving the meaning of work?

BouclePro’s answer can be summarized in one sentence: the unit is not the profile, the unit is the Loop.

A Loop is not a feed. It does not merely circulate messages. It keeps context, organizes memory, makes contributions visible, connects people, welcomes unclear requests, and turns individual learning into common resources.

In 1996, CyberWorkers said: work can leave the office.

In 2026, BouclePro adds: work must not leave the collective.

The future of work will not be decided only by the power of AI models. It will be decided by our ability to create spaces where humans can still ask, help, learn, decide, transmit, and keep the thread.

Sources

Weak ties, human signals, and language

• Mark Granovetter, The Strength of Weak Ties

• MIT Initiative on the Digital Economy, New Study Proves That Weak Ties Have Strong Employment Value

• CNRS, Can algorithms detect humor, emotions, or hate speech?

Silence, conversational AI, and human extraction

• Korben, Meta employees do not want to install a logger on their PC to train AI

• ELIZA effect

• Jaron Lanier, Digital pioneer Jaron Lanier on the dangers of free online culture

Dialogue, questioning, and collective methods

• Socrates, Stanford Encyclopedia of Philosophy

• Plato on Knowledge in the Theaetetus, Stanford Encyclopedia of Philosophy

• David Bohm, On Dialogue

• Edward de Bono, Six Thinking Hats

• Edward de Bono, Lateral Thinking

• F. David Peat, Gentle Action

• Robert S. Root-Bernstein and Michele Root-Bernstein, Sparks of Genius

• Climate Fresk

Learning, communities, and social health

• WHO, Commission on Social Connection

• WHO, Depression

• Julianne Holt-Lunstad et al., Social Relationships and Mortality Risk

• Vivek Murthy, Together

• The Social Cure: Identity, Health and Well-Being

• Jean Lave and Etienne Wenger, Situated Learning

• Malcolm Knowles, informal adult education, self-direction and andragogy

Human-in-the-loop, human oversight, and AI feedback

• Label Studio

• Argilla

• InstructLab

• CNIL, Using an AI system in production

• European AI Act, Article 14

Platform work, commons, and recognition of contributions

• ILO, The role of digital labour platforms in transforming the world of work

• Antonio Casilli, Waiting for Robots

• Open Source Initiative, The Open Source Definition

• Peer Production License

• Platform Cooperativism Consortium

• Zindi

• Open AIR

• ReFFAO

• RFFLabs

Art, science, telepresence, and caravanserai

• Roger F. Malina, CaravanserAI

• Eduardo Kac, Rara Avis

• Leonardo / ISAST, Award Recipients

• Leonardo

• MIT Press, Leonardo Journal

• OLATS

Workplace, organizational memory, and critique of the feed

• Meta Workplace

• TechCrunch, Meta is shutting down Workplace

• Mimecast, Meta Workplace shutting down

• Interact, Meta to shut down Workplace platform

• Workplace, encyclopedic entry

• Reworked, Meta shutters Workplace

• IT Pro, Meta’s Workplace was always destined to fail

AI agents, governance, and AI Shepherding

• NIST, AI Risk Management Framework: Generative AI Profile

• European Union, AI Act, Regulation (EU) 2024/1689

• UNESCO, Recommendation on the Ethics of Artificial Intelligence

• Stanford HAI, What is Human-Centered AI?

• Anthropic, Building Effective AI Agents

• OpenAI, Building agents

• OpenAI, A practical guide to building agents

• OpenAI, The shift to agentic AI: evidence from Codex

• Anthropic, Measuring AI agent autonomy in practice

• OECD, Five principles for the effective use of AI in vocational education and training development

• Eren Bilen and Justine Hervé, When AI Gives Bad Advice

• OpenAI, Self-Evolving Agents

• Limits of Safe AI Deployment: Differentiating Oversight and Control

• Auditing and Controlling AI Agent Actions in Spreadsheets

• California Management Review, Rethinking AI Agents

• ILO, Disruption without Dividend?

• Fairwork, GPAI Fairwork Ratings Report 2024/5

• Fei-Fei Li, A Human-Centered Approach to the AI Revolution

• Stanford HAI, AI in the Loop: Humans Must Remain in Charge

• Yoshua Bengio, Scientist AI

• Iason Gabriel et al., We need a new ethics for a world of AI agents