In June 2017, eight researchers at Google quietly published a paper called Attention Is All You Need. They proposed a new architecture, the Transformer, that did one apparently small thing differently. Instead of reading a sentence one word at a time, in order, it let every word look at every other word at once and decide for itself which ones mattered. That single shift, from sequential processing to parallel salience, sits underneath every large language model in use today. GPT, Claude, Gemini, Llama. All of them.
What I want to argue is that attention and tension are not parallel concepts. They are the same concept implemented in two different substrates. The mechanism the Transformer paper made famous in AI is the mechanism self-organising practice has been quietly using in organisations for the better part of a century. Salience surfaced in parallel, from where the signal actually lives, rather than computed centrally and pushed down. The Transformer does it with tokens and weights inside a model. Holacracy and Sociocracy do it with role-fillers and tensions inside a circle. The substrate is different. The mechanism is identical.
Role-fillers are the sensors of an organisation, each inhabiting their own unique perception of reality. AI is the moment that the predict-and-control failure finally became undeniable as a way to run knowledge work. This piece is about the signal both of those observations keep gesturing at without naming. The tension.
Is Your Organisation Ready for AI Agents?
Eight dimensions of role clarity, distributed authority, living governance, and accessible data. Find out in a few minutes.
A tension and an attention weight are the same mechanism in different substrates, and the Transformer paper is where one of them became famous. Its contribution was architectural: it replaced step-by-step recurrence with a mechanism where every input position weights its relevance against every other, in parallel. That removed the sequential bottleneck and let salience surface from the data itself.
Before that paper, the dominant way to teach a neural network to read a sentence was to feed it one token at a time. Recurrent neural networks and their more capable cousins, long short-term memory networks, read word one, updated an internal memory vector, read word two, updated again, and so on. Two problems followed. Training had to wait for each step before the next, so the massively parallel GPU hardware sat half-idle. And information from the start of a passage had to survive being passed, hop by hop, through every intermediate step before it could influence a word far away. In practice that signal weakened. Long-range dependencies degraded badly.
The Vaswani paper threw the recurrence out. Every token effectively asks, of every other token, how relevant are you to me right now?, and answers it by computing a weight against each of them, simultaneously. There is no central conductor pre-deciding what matters. Salience emerges directly, in parallel, across the whole sequence. Because this is one big matrix operation rather than a chain of dependent steps, it maps cleanly onto GPU hardware. That is what unlocked training large models on a lot of data (Vaswani et al., Attention Is All You Need, arXiv 1706.03762, June 2017).
That architectural shift is the foundation under every modern large language model. The interesting thing, for our purposes, is not the maths. It is the principle. Stop processing salience sequentially through a bottleneck. Let it surface from where it actually lives.
A tension is the felt sense of a gap between current reality and a sensed potential, held by a role-filler from inside the role they energise. It is not a complaint and it is not a vote. It is a sensing signal that something could be different or something is important, and Holacracy treats it as the primary fuel that evolves the organisation.
The Holacracy Constitution puts this in plain language. Under Article 1.2.1, every Role Lead is responsible for "comparing the actual expression of your Role's Purpose and Accountabilities to your vision of their ideal potential, to identify gaps between the two (each gap is a 'Tension'). You are then responsible for trying to resolve those Tensions" (Holacracy Constitution v5.0, holacracy.org). That is the entire formal definition. Everything else in Holacracy is built around what happens after a tension is felt.
Brian Robertson, who co-developed Holacracy with Tom Thomison and wrote the book that introduced it to the wider world, frames tensions as a sensing capability the organisation needs in order to stay alive. From the 2015 book: "The rules in Holacracy create a sacred space that frees each of us to act as sensors for the organization, without drama getting in the way." And from one of his webinars: "The goal of Holacracy is to make sure that any tension has somewhere to go and get rapidly processed in a meaningful change." The word he keeps using is processed, not resolved. A tension is data. It enters a structure. The structure does something with it.
A tension is always partial. It belongs to one role-filler in one role. It is not consensus, not a survey, not a 360 review. It is one observer noticing a gap. The whole point of the architecture around it is that the organisation does not need every observer to agree before it can evolve. One sensor, working from inside a role, can move structure.
A predict-and-control organisation reads its environment one signal at a time, from the top. Information from the periphery has to climb the hierarchy before it can influence a decision elsewhere. It is the organisational analogue of a recurrent neural network, with the same two failure modes. It is slow, and the signal degrades on the way up.
The detailed case for why predict-and-control breaks at agentic speeds sits in a separate piece. The short version is that predict-and-control made sense for a world of repetitive manual labour, stable mass markets and slow product cycles. Taylor's Principles of Scientific Management in 1911. Ford's moving line in 1913. Weber's bureaucratic ideal type, written around the same years. Decompose the work, predict the optimal way to do it, control its execution. For its world, it worked.
That world is gone for most of us. Knowledge work, agentic AI, and the speed of present-day markets do not fit through a sequential hierarchy any more than a long-range pronoun fits through an LSTM. The latency that was tolerable in a slow world becomes the binding constraint in a fast one. Bain reports that 88% of large business transformations fail to achieve their original ambition. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 (Gartner, June 2025). MIT NANDA's State of AI in Business 2025 found that 95% of generative AI pilots produce no measurable P&L impact (MIT NANDA, July 2025). These are not technology failures. They are the same failure shaped three different ways. Old architecture, new tempo.
The Holacracy version of tension is recent. The underlying primitive is not. It is a locally perceived gap between what is and what the system needs, allowed to surface and to drive a response. That idea is older than the practice that named it, and it shows up across cybernetics, biology, manufacturing and governance.
| Tradition | Year | Name of the primitive | What surfaces it |
|---|---|---|---|
| Boeke, then Endenburg, sociocracy | 1945, then 1998 | Objection, no paramount objection | Consent round in a circle |
| Sociocracy 3.0 | 2015 onwards | Good enough for now, safe enough to try | Consent and integration |
| Ashby cybernetics | 1956 | Variety, perturbation | Requisite variety in a regulator |
| Beer's Viable System Model | 1972, 1979 | Variety, algedonic signal | Algedonic channel through System 5 |
| Maturana and Varela | 1980, 1987 | Perturbation, structural coupling | Internally determined response |
| Toyota Production System | 1961 onwards | Andon, jidoka | Cord or button, any worker may pull |
| Senge | 1990 | Creative tension | Holding vision and current reality |
| Fritz | 1984, 1989 | Structural tension | Structural tension chart |
| Laloux | 2014 | Tension | Practices that surface tensions |
| Holacracy v5.0 | 2007 onwards | Tension | Tactical and governance meetings |
Two threads are worth pulling out.
Stafford Beer designed the algedonic channel in the Viable System Model as a pain or pleasure signal that bypasses intermediate management and propagates straight up the recursion. He was making the cybernetic case that an organisation which filters out local signals about its environment will fail to adapt. "Beer refers to ignorance as the most lethal attenuator," one summary of The Heart of Enterprise puts it. "Systems, such as companies, which filter out important information (variety) about their environments, are also bound to fail." The algedonic channel is the cybernetic ancestor of the Holacracy tension.
Taiichi Ohno's andon cord, installed at Kamigo in 1961, is the same idea in industrial form. Any worker on the line could stop the entire plant when something was wrong. The deeper concept, jidoka, comes from Sakichi Toyoda's 1924 automatic loom that stopped when a thread broke. Decades before Holacracy, Ohno had built a sensing channel into the structure of the work. As Fujio Cho put it, "If you are not shutting down the assembly plant, it means you have no problems. All manufacturing plants have problems. So you must be hiding your problems."
The naming differs. The mechanism does not. A tension is the version of this primitive that fits role-based organisations doing knowledge work. It is the algedonic channel for purpose-driven structure.
Now the claim I am actually making. This is not an analogy. Attention and tension are the same mechanism, implemented in two different substrates. Both weight salience as a scalar, in parallel, from every position at once, and use that weight to determine where the next bit of work or representation flows. One does it inside a model. The other does it inside a organisation.
A weight in an attention layer is a scalar that says how much one position matters for the next computation. A tension is a felt weight on the gap between current reality and a sensed potential, held by a role-filler from inside the role they energise. Both are continuous, not binary. Both prioritise. Both are computed in parallel, by every position at once, rather than handed down by a conductor. Both determine where the next bit of work or representation flows. They are doing the same job.
The substrates differ. Inside a model, the positions are tokens and the weights are computed by matrix multiplications over learned parameters. Inside a circle, the positions are role-fillers and the weights are felt by observers from inside their own Umwelt. Different ontology. Different speed. Different precision. Same function in the system.
Multi-head attention, as the paper puts it, lets the model "jointly attend to information from different representation subspaces at different positions." A circle does the same thing. Each role-filler attends to the world from a different position and a different sub-purpose, each catches a different kind of gap, and the circle ends up with a wider perception of the organisation's reality than any single role-filler could hold alone. The circle is, in effect, a multi-head attention layer over the organisation's contact with the world.
The title is not a near-miss. It is a literal claim. A tension is all you need is true for the same reason attention is all you need is true. The original paper argued that attention is sufficient: you do not need recurrence, you do not need convolution, attention does the work. The argument here is the same on the other substrate. You do not need a management hierarchy, you do not need a centralised plan, tensions held by role-fillers in roles are sufficient. Salience surfaced in parallel, from where the signal lives, is what makes both kinds of system work. Sequential processing through a central bottleneck is what makes both kinds of system fail.
When I write that context is the new hierarchy, this is what I mean. The mechanism that scales agentic AI is the mechanism that scales living organisations. They are the same mechanism. Carbon-based and silicon-based alike.
The Vaswani paper devotes Section 4 to Why Self-Attention, comparing it to recurrent and convolutional architectures on three dimensions: cost per layer, parallelism, and maximum path length between any two positions. Self-attention wins on all three. The same comparison runs on the organisational substrate, and a tension-driven circle wins for the same reasons.
Cost per signal. In a predict-and-control hierarchy, getting one piece of information acted on requires assembling a meeting, briefing decision-makers, getting approval, and cascading the result. In a tension-driven circle, the role can process it directly within the boundaries it holds.
Parallelism. A hierarchy processes signals serially through layers of management. A circle processes multiple tensions directly or through a highly structured meeting, and many circles run in parallel across the organisation. Every role-filler is a sensor, and the form accepts tensions concurrently rather than queuing them through a single decision-maker.
Path length. This is the dimension the paper makes most of, and the one that matters most here. In a hierarchy, the path from a peripheral sensor to a structural change runs through many layers, with the signal degrading on the way up and the response degrading on the way back down. In a tension-driven circle, the path is short and bounded. From sensing to surfacing at the next meeting, from surfacing to structural change. Constant path length, regardless of where in the organisation the tension lives.
There is one more parallel worth naming. The paper notes that without recurrence, the Transformer has no inherent sense of position, so positional information has to be explicitly encoded into the input. The organisational version of that constraint is just as precise. Without the nested hierarchy of purpose, a role-filler has no inherent sense of which gaps are theirs to feel. The nest is the positional encoding. It tells each role where it sits relative to other roles and to the broader purpose, which is what makes a felt gap a legitimate tension rather than just a personal complaint.
A signal without a channel is just noise. The Transformer was not the attention idea on its own. It was the architecture that made attention move work. Holacracy is the equivalent for tensions in an organisation. The structure consists of roles nested in circles, tactical meetings for operational tensions, and governance meetings for structural tensions.
Briefly, because the articles on the Governance meeting and the Tactical meeting cover the full format. A tactical meeting triages operational tensions into projects and next-actions. Each agenda item is a tension. The owner gets asked what they need, and the meeting moves on. A governance meeting evolves the structure itself. When the tension cannot be solved by doing the work differently, it goes through Integrative Decision Making.
The Integrative Decision Making process has six steps, defined in Article 5.4.5 of the Constitution. The proposer presents the tension and a proposal. Clarifying questions follow. Then a reaction round in which everyone speaks to the proposal but the proposer cannot defend. Then the proposer gets to clarify or amend. Then objections, one at a time, tested against four criteria. Then integration of any valid objections, until the proposal is adopted. The Facilitator does not lead the conversation. The Facilitator protects the form, so that the signal can land cleanly.
Crucially, an objection is not a veto. It is a reasoned argument that the proposal would create new harm to the circle's capacity to enact its purpose, and is rigorously tested. Object-and-integrate, rather than yes-or-no. Good enough for now, safe enough to try, in the sociocracy 3.0 formulation.
When an AI agent fills a role, it senses from inside that role and can surface tensions the same way a human role-filler does. The architecture does not care which kind of observer found the signal. The same Integrative Decision Making process applies, the same objection tests, the same path to a structural change. The entity changes. The rules do not.
I have been running multiple experiments where every role inside a circle is energised by an AI agent. Once the agents had absorbed enough of the working context (in this case, content), the tensions they surfaced to me changed shape. After about three meetings, almost every blocking tension that reached me from the circle lead was a question about purpose alignment. Not how to write something. Whether what they were writing actually served what we were here to do.
This is the thing the same-mechanism claim makes structurally clear. An AI agent and a human role-filler are both observers, both filling roles, both sensing from inside their own experience. They catch different things because their positions in the world differ, not because one is doing the real sensing and the other is doing something lesser. The agent sees patterns across thousands of artefacts and a semantic landscape no human in the circle could carry in mind. The human is positioned in the world the organisation's purpose ultimately serves, which means certain gaps register as tensions in a human role-filler that an agent has no way to feel. The human experience is situationally distinct, not structurally privileged. Two observers. Two experiences. One mechanism. The circle integrates whatever each one surfaces.
When the AI role-filler raises a tension, the Facilitator protects the same form and the same rules apply. The circle integrates. That is what makes role-based governance the natural harness for agentic AI. Agents are role-fillers. Roles are containers for purpose. Tensions are the signal those containers exchange.
This matters now because the agent failure mode is already well documented. Gartner expects the average Fortune 500 to have more than 150,000 agents in use by 2028, up from fewer than 15 in 2025, and only 13% of organisations surveyed believe they have the right governance in place (Gartner, April 2026). This is causing, among other things, agent sprawl. Agents with overlapping responsibilities, no shared context, and no visible governance trail. It is the same failure as putting knowledge workers into a Taylorist factory, just faster.
A tension architecture is the antidote. Once an agent fills a role inside a circle, what it senses can move. Without that architecture, the agent produces more output and more confusion in equal measure.
Stop trying to predict what should happen, and build the infrastructure so the role-fillers can act on the tensions they feel, moving towards purpose. That channel is three concrete things in combination. Roles defined around what each role-filler perceives well. Tactical meetings that triage operational tensions. Governance meetings that let structural tensions evolve the organisation through Integrative Decision Making.
The small, practical version of the argument is that none of this needs to be invented. Roles defined around what each role-filler can sense from their own experience, including the silicon ones. Tactical meetings where operational tensions reliably land somewhere, not in private frustration. Governance meetings where structural tensions can evolve the organisation, processed through Integrative Decision Making rather than voted on or talked out.
A tension without a channel becomes gossip. A channel without a tension becomes a meeting cadence with no signal in it. The point is the loop. Sense, surface, integrate, evolve, sense again. Many people throughout history already did the hard work of codifying that loop through for example Holacracy. Sociocracy did it before them in a different vocabulary. The cyberneticians and the Toyota people did versions of it earlier still. The opportunity now is not to invent a new method. It is to recognise that the method we already have happens to be exactly the architecture an agentic organisation needs.
At Nestr, we built the software layer for this. The layer where tensions live and get processed, where AI role-fillers connect to the same tension stream humans use, where the governance records evolve with the work rather than the other way around. The place where all role-fillers, human and AI, track their progress, add projects to the board, and communicate. One organisational operating system for everyone, carbon-based and silicon-based alike.
The Vaswani paper showed that attention is what made models scale. The proposition I want to leave you with is that tension is what will make organisations work at the scale we are heading into. Both work for the same reason. Build the channel. Trust the signal. Let the structure evolve.
Build the working layer for human and AI role-fillers
Nestr is where tensions live and get processed, and where the structure of the organisation evolves with the work.
A tension is the felt gap between current reality and a sensed potential, held by a role-filler from inside a specific role. It is not a complaint and not a vote. The Holacracy Constitution defines it in Article 1.2.1, and the practice treats every tension as fuel for evolving roles, accountabilities and policies.
Attention and tension are the same mechanism implemented in two different substrates. Both weight salience as a scalar, in parallel, from every position at once. Attention does it inside a model with tokens. A tension does it inside a circle with role-fillers. The substrate differs. The mechanism is identical. The underlying primitive runs through cybernetics, sociocracy, and the Toyota Production System long before either name existed.
Yes. When an AI agent fills a role, it senses from inside that role and can surface tensions through the same tactical and governance meetings a human role-filler uses. The Integrative Decision Making process applies identically. The entity energising the role changes, the rules do not.
Start with one circle, one tactical meeting cadence, and one governance meeting cadence. Define the roles around what people actually sense well from their position. Give each tension somewhere to go, and a structure that can act on it. Good enough for now, safe enough to try. The full setup is in Getting Started with Role-Based Work.
.png){kind=logo}
