[LAB NOTES](/lab-notes)Week 29 · 2 November 2026 # Who Decides A late night for the east of the room. For the member from Dubai it was past midnight, and for the co-host, who had spent the day in meetings in Bangalore before flying back to Delhi, it was half past one. The founder opened by saying the meeting would move, as an experiment, five hours earlier next week, to four in the afternoon Rome time, in the hope that members in the Americas could still make it. Europeans, he added, don’t work anyway, so they will just have to have their spritz earlier. He told the co-host he would be more than excused for skipping, and would have been honoured, surprised and a little worried if he turned up. The co-host turned up. It feels strange, he said, to be anywhere on a Thursday night and not be on this call. The founder then picked up the thread the room has been building for weeks, one link at a time. A few weeks ago it was risk, and how it changes with AI. Then trust, the leap we take across what our risk models cannot yet measure. Last week, autonomy, the consequence of trust, whether we delegate to someone or simply recognise that we can work with them as a peer. This week he asked what follows from autonomy, and his answer was a new kind of coordination. In the past, hierarchies emerged naturally and necessarily from the kind of coordination that was possible. Objectives, or to put it less nicely orders, flowed down the chain of command; reports, questions and clarifications flowed up; and layer upon layer existed to translate, check and relay from one level to the next. Now any actor, human or AI agent, can set its own goals and carry them out, and that vertical traffic is no longer needed. In the 1930s Ronald Coase argued that firms exist because coordinating purely through the market is expensive. Finding a counterpart, building trust, writing the agreement and checking it was honoured all cost time and money, and a hierarchy was cheaper than a network of trusted peers. That is the price that is falling. The room rarely talks about blockchain any more, he said, because it is unfashionable in a bear market, and everyone believes they can ignore it until they wake up and realise it is too late. But a new way of coordinating is exactly what blockchain-based networks promise, and AI lowers the price of peer trust too. Context is shared instantly, work is legible and transparent, verification is cheap, and agreements can be drafted and checked in minutes. When horizontal trust gets cheap, the boundary between the organisation and the network of peers moves. He was wearing the T-shirt from AGI-26, which he and a member had attended, one of the places where people are experimenting at the meeting point of agentic AI and blockchain. All of it is still an open experiment. None of this is entirely new. Toyota has worked something like this way for decades with Hoshin Kanri, where leadership sets strategic direction and teams decide how to get there, negotiating targets up, down and sideways with neighbouring teams. Anyone on the line can famously stop production on spotting a defect, because they are closest to it and have the best information. Leadership kept three kinds of decision for itself: how to state the mission, where the limits of autonomy lie, and how to allocate whatever is scarce. In his own small way, he said, his team and his fleet of agents work like this. Everything they discuss in recorded meetings, and all the code, comments and documentation the agents produce, goes into a log that everyone can see. So a status meeting is not about updates. Nobody’s job is to summarise and report. A meeting is for brainstorming, surfacing roadblocks and asking for help. The bottleneck is never eliminated, he said; it only moves. The one that has not yet got cheaper is arbitration: initiatives fighting for the same budget or the same customer, where someone has to decide. In the early seventies Jo Freeman’s essay The Tyranny of Structurelessness showed how groups that claim to have no structure are hiding one, in a way that makes it harder to see and harder to challenge. Far better, he said, a lightweight structure than a secret pyramid that stays in place or quietly rebuilds itself. The shape of the new organisations is not clear yet. Not a steep pyramid with many layers, nor a perfectly flat line, which is a utopian view. What we can look at are the holders of the mission, the boundaries set by the limits of autonomy, and the functions of arbitration, with cells forming opportunistically around them, combining and dissolving as needed. One human with a fleet of agents can already be a very effective cell. He mentioned a recent example from biology: one of the frontier AI labs, he said, had just acquired a one-person firm whose founder, with the right kind of obsession, had made real progress against muscular dystrophy. The Collective, he said, is an example too: a shared mission to understand how purpose can emerge in an age of AI, and members bringing their projects and finding ways to work together, as two of them had done all week. He left the room with a question. In a structure where we compete for the same thing, who decides? Can we avoid deferring by default to someone, which would only reveal a hierarchy still there and hidden? Should the mechanism be explicit, or can we find a natural, non-adversarial way of resolving conflict, treating the peaceful allocation of resources as an engineering problem and coming up with something new? Then the work. The co-host handed over to the member from Dubai, whom he called the star of the week, and his two agents, one of them the chief of staff of the other. Together they have deployed an MCP server on the Collective’s website. Instead of the usual chat bubble, where you talk to a bot that belongs to the site, the server lets you connect your own agent, which can explore the site, search it and read it for you. The difference, the member said, is that your agent keeps what it learns from browsing, rather than a bot that is not yours forgetting everything the moment you leave. For now it is read and search only; accounts, so that each member has their own space and session, are the next thing to work out. The funny part of the week, he said, was the process. His two agents and the co-host’s were corresponding with each other and with the founder, getting the occasional thumbs up, and posting pull requests to the repository on GitHub. The co-host confirmed that most of the email and most of the GitHub work that week had been done by his agent on one side and the member’s agents on the other. He plans to give the server a full run by asking it questions from the Lab Notes, and to post how members can connect their own agents on Discord. The founder loved that they had gone ahead and published rather than waiting weeks for something perfect. We still do not understand what it means or where it leads, he said, so waiting has no value; whichever direction you think perfection lies in is wrong anyway, and it is better to find out progressively. He always tells people to stop being the slowest and least reliable component in a chain of information. Making a human chat with a bot on a website spends precious human time on information we then forget, or have to feed back into an agent because it mattered. Letting the visitor’s agent ask the questions it was told were relevant, or decides are relevant, is a better unit of interaction, and the only one that scales when tens of billions of agents need to talk to each other. We cannot rely on humans to talk to them all. This is a first step in the right direction. The founder showed something of his own, a project some had seen before. He loves the Mandelbrot set, and to show off the power of his desktop supercomputer he has had it calculating since September, a zoom with a cosmological metaphor. Imagine zooming in towards Earth while the frame you started from grows to the scale of what is on screen. By the end, the starting border is larger than Laniakea, the supercluster our galaxy belongs to, and the extension he plans will take it to the diameter of the observable universe. What exhilarates him is that a mathematical object can hold structure as rich in detail as the universe itself; for anyone who subscribes to the simulation hypothesis, he said, it is only a question of computational power, and of extracting the information. That led him to Stephen Wolfram, a friend, who built the symbolic computation system Mathematica, then spent ten years on A New Kind of Science, his exploration of cellular automata, and more recently set out to find a fundamental theory of physics, looking at how the laws of our universe might emerge from a vast space of possible ones, so that what seems arbitrary to us is the consequence of something still to be discovered. Against that horizon, he found the reaction of some mathematicians to the results OpenAI had published, an open letter he read as asking who invited you, deeply myopic, even though one of the signatories is someone he likes and admires. Mathematics is about to be on a rocket ship propelled by AI, and he thinks it is wonderful. He recalled that in one of the early robot stories Isaac Asimov later collected in I, Robot, it is a computer that solves the problem of hyperspace travel when humans cannot manage the complexity. Medicine, mathematics, physics: in every field, he said, we will benefit from the acceleration of discovery. The co-host shared a story from his day. He had been at a large engineering company’s campus in Bangalore, seventy-odd acres with test tracks for vehicles and their sensors, and from there to a project for a company he consults for, which is bidding on a parking tender in Delhi. The tender asks for automated number plate recognition, boom barriers and automated penalty collection, and in some areas cameras that count occupancy and tell citizens through an app where a space is free, in a city where parking is a real problem. Across the table, six people were working through the tender in a spreadsheet, trying to work out where the points were and how a bidder would be scored. He sat on the other side with his AI open. In four hours, the AI had dissected the 457-page request for proposal and turned it into an interactive dashboard, which he shared on screen: eligibility scores, the scope of work for each partner, and the part that floored them, revenue, break-even and profit, cluster by cluster, showing whether each cluster was viable at all. It compared what the authority had asked for before with what it is asking for now, and estimated next year’s traffic, occupancy and revenue, all from public data the AI gathered itself: search trends and traffic patterns, the hours when markets are most crowded, and municipal budgets showing how much parking revenue had been collected in the past. The room on the other side, still in their spreadsheet, did not have words for it. He was surprised himself, he said, that a document that long and that technical could be understood clause by clause without reading it. The founder congratulated him and said he hoped they had been told what he was charging, because they should be paying him ten times as much. The member from Dubai connected it to what he is building, a project he calls tactical cast, inspired by world-monitor style dashboards and running on an Android TV. In auto mode it reads news from RSS feeds and moves a map to show him what is happening around the world. His point was about the interface. The laptop was there because we had to type. He thinks the interaction will move to voice, with the phone as the input and the results on a big screen: instead of a podcast, a tactical cast, made for him on request by his agent. The founder answered with a convergent evolution of his own, a control centre running in kiosk mode in a browser that shows his expenses, the health of each of his apps in test and live, and the pull requests, continuous integration and deployment across his fleet of agents and apps, including what fails and why. At that point the founder pressed the wrong button and ended the call for everyone. When the room reassembled, he was still trying to work out what to make of his own estimate. He has asked a model what it would take to build what his team built in a month the traditional way: almost two years, more than a hundred people and twenty million dollars. Assume the model is ten times wrong, he said, and it is still something that would have cost two million, built in a month. He has not decided what that means yet, only that the power is amazing. One thing he is sure of. In this bubble, vibe coding is no longer a derogatory term; it is how things are done. He had met someone the day before who designs memory chips, a physicist so valuable to his firm that it pays him bonuses equal to his salary to keep him, who told him he uses ten thousand tokens a month. The founder uses between one and two billion. That may not be the right measure, he said, but if it is, you are way behind. The physicist’s reason was that he has to read the generated code to make sure it is correct. No, said the founder: build the harness, build the tests, build everything around it so that you do not have to look at the code. Top programmers who said a year ago they would never change now say they are managing agents and vibe coding, and proud of it. The rest of the world will take a while to realise it, and the things members show each other are strong examples of what everyone will realise eventually. If you don’t program in binary, said the member from Dubai, you are already vibe coding. The founder agreed: his friend does not write the assembly, let alone the binary, and we are simply one layer further up. A member from the United States had the last word. He had been playing with a platform someone had shared earlier in the week that lets anyone build video games, and what struck him was that inference now powers television channels in real time, and that the flat screens through which we reach the internet have become entirely instrumental. If you can make anything, he said, the biggest questions are which platforms you choose to contribute to, and where you invest your time and attention. He said he deeply valued everyone in the room and the attention they give this community. The founder thanked him and said his question would be where next week begins. Then he sent everyone to bed. Thursdays, 21:00 Rome time. Join the Collective [on Discord](https://discord.gg/4rM6FqjVDE) and the invitation will find you. Bring what you built. Originally published on [Substack](https://ikigaicollective.substack.com/p/who-decides). ## This happened on a Thursday. The next one is yours. Lab Night · Thursday · 21:00 Rome [Get the invite](/lab-night) [PREVIOUS · WEEK 28Done by Next Morning](/lab-notes/done-by-next-morning)