Harnesses, Bodies and Deaths: Notes from an Evening

What turns a pile of simple parts into something that makes sense of the world? A neuron on a chip, an ant in a colony and a language model inside an agent swarm all face the same problem: each follows local rules, yet something coherent is expected to appear at the level of the whole. This session had no talk. It was an open evening of conversation between researchers, engineers and founders, and it kept returning to one question from different sides: what does a system need beyond raw intelligence in order to adapt? The answers that emerged were a harness, a body, consequences, and, surprisingly, a way to die.

Two ways to study complexity

The evening opened with a distinction between two traditions in complexity science. The first is reductive. It compresses messy phenomena into short, interpretable equations in which every parameter means something. The classic example is the scaling law linking heart rate, body size and lifespan: across mammals, the number of heartbeats per lifetime is roughly constant. A mouse's heart races and it lives two or three years; a whale's heart beats slowly and it lives for decades.

The second tradition is emergent. It starts from a vague intuition, writes rules for individual agents, and tunes the parameters to see what appears. One participant described an agent-based model of how dominance hierarchies form, built by adjusting mating preferences and evolutionary pressures until stable hierarchies appeared. Spiking neural networks belong to the same family: simple units, local rules, and the hope that interesting behaviour arises from their interaction.

The room leaned towards the second tradition. It does not dissect the world; it adds to it. Its difficulty is also clear. Open-ended emergence is hard to fund outside academia, because no one can say in advance what will come out. One answer was to anchor it to a concrete problem, such as human disease, where there is no shortage of complexity to tackle.

What a single cell computes

The conversation then moved to biology on silicon. Neurons grown on a microchip start to respond to the signals they receive and to organise around them. At the level of a single cell, the working hypothesis is prediction error: the cell acts to anticipate what comes next and adjusts its internal model when it is wrong. That raised a question no one could answer: how many transistors does it take to match one neuron's capacity to minimise prediction error?

A related question was whether individual neurons are computationally reducible, meaning whether a set of rules could predict exactly what a neuron will do given its inputs. With enough data, perhaps. The problem is efficiency. Even the most advanced models of a fly brain need a supercomputer to run, and the widely shared fly connectome maps the wiring without capturing how the circuit actually fires. Knowing the structure is not the same as knowing the dynamics.

This led to a sharper disagreement about hype. One view was that there are no longer filters on claims, so demonstrations of "thinking" brains in a dish sell well to people who cannot check them. The other view was that the filter still exists; it is simply slow. The proof is in the pudding: what does not work eventually fails, although it can take a long time and a lot of money on the way down.

Creativity needs a harness

Some of the most concrete findings of the evening came from people who build agent systems daily. Asked to generate synthetic conversations for training a classifier, a frontier language model began looping after twenty or thirty examples. Instructing it to "be creative" changed nothing. Telling it to look at earlier examples and avoid them made things worse: whatever sits in the context window gets reproduced. It is the pink elephant problem. Being told not to think of something puts it firmly in mind.

What worked was a deterministic randomness harness. Plain code rolls virtual dice across many dimensions (why a visitor came to the theatre, how their day went, what they ate) and hands the model a fresh combination each time. The result was tens of thousands of distinct scenarios. Deterministic code, used well, makes a model's output effectively non-deterministic.

A second observation was stranger. Telling a model it will play a role "for ten rounds", even when it will only play one, makes it perform better. People behave the same way: we put more effort into a task we will own for a month than into one we will do once. Models, it seems, are overconfident about their own creativity and only become creative when a structure forces them to. The room suspected the same is true of humans.

One participant put this as a general claim: intelligence is an architecture, not a model. A static model cannot be intelligent on its own. The harness gives it scenarios, objectives and time horizons; in effect, it gives it something close to experience.

Swarms, hierarchies and the orchestrator

What happens when many agents work together? The evening produced a small catalogue of experiments. An agent system organised as a rigid corporate hierarchy, with a chief executive agent, a technical lead and so on, produced extreme bureaucracy: large amounts of wasted compute and no better results than a single background sub-agent. At the other extreme, a sandboxed swarm given a budget and no instructions looped endlessly. Its internal bureaucracy grew exponentially, and it rediscovered facts that earlier sessions had already found. Left long enough, such swarms invent message boards and teams, which amount to the sub-agent structures they could have been given from the start. Even light orchestration gave roughly a tenfold efficiency gain over the free swarm.

What seems to work today is simple: a coordinating agent, a few "lanes" beneath it, and sub-agents spawned from those lanes. The human remains the bottleneck, deciding what to do and in which direction to steer.

This fed into a debate about scale versus quality. Would you rather have ten thousand interns in a sports hall or a hundred experienced engineers? The engineers win, because they know where to look and whom to call. A slightly more capable agent makes one fewer mistake, and one mistake can send an entire swarm in the wrong direction. The counterargument came quickly: in a genuinely new paradigm, where no one knows the answer, volume matters more than experience. A third option drew the most agreement: a hundred experts from unrelated fields (musicians, firefighters, doctors) plus a handful who know the domain well. That is brute-force search with a better distribution of priors.

The same logic undermined a popular idea: routing each request to the cheapest model that can handle it. Deciding where a task should go requires understanding the task. The routing decision needs the smartest model in the system, not the cheapest.

One proposal remained open. The orchestration layer of an agent system may be small enough to model probabilistically. Its states are tool calls, queries and context injections. Logging them and building a decision tree of likely next actions could make it possible to work backwards from a goal. Whether that would automate the human orchestrator, or merely describe them, remains to be tested.

Where the giants cannot reach

A recurring concern was the gravity of the large AI laboratories. As they move into fields such as biology, what is left for small teams? One heuristic was to ask where the incumbents are spoiled: where they have so much of something that they never need to economise on it. Sample-efficient active learning for experimental design is one candidate, since a large lab can always buy more annotation.

The pushback was that the laboratories are constrained almost everywhere, because demand outstrips what they can supply. The one resource they have in abundance is attention and credibility. That makes the heuristic harder to apply than it first appears.

A second tension concerned strategy. Some founders choose to supply data to the laboratories rather than compete on intelligence. To some in the room, this felt defeatist: if the aim is to cure disease, why hope someone else will do it? A more constructive answer was that intelligence may already be sufficient for many problems. The bottleneck is the glue between that intelligence and the real world: the laboratory, the clinic, the physical process. Self-driving laboratories were one example. Much of the biological literature is shaped by researchers who built careers on a particular niche and are subconsciously pushed to confirm it, for instance by adjusting a microscope's sensitivity until the expected result appears. This is not malice. It is the same bias a language model shows when it reproduces what sits in its context.

The image that stayed was gravitational. The laboratories are planets. Come too close and you are pulled in. But most of the space of possible intelligences lies far from them, and much of it is unexplored. Their pull comes from human incentives, not from anything the models themselves want.

Consequences

One of the most unusual projects discussed was a small wooden box with a vibration motor, a knocker and a camera. Drop it, and it becomes "scared": the temperature of its neural network shifts, and it responds differently. The next step is to make its learning irreversible, so that every experience permanently rewrites its weights, and to make those weights impossible to copy. Dropping it, or treating it kindly, would then have consequences in the full sense: something happens that can never happen again.

The inspiration was an injured pigeon, nursed back to health over weeks. It had very little intelligence, yet the bond between it and its carer was unmistakable. That observation led to a system modelled on the brain's older, non-reasoning layer. Its first version kept four "buckets" of neuromodulators that filled and drained like leaky containers. Fear added adrenaline and a little dopamine, because fear is also exciting. A camera read the user's face and the conversation transcript, and the bucket levels modulated the language model's behaviour.

The key finding was that numbers mean nothing to a language model; stories do. Telling the model "anger: 0.8" changed nothing. Telling it "you feel tense, your replies are shorter and sharper" worked. The old feeling-story also had to be removed from the context at every turn, or it would accumulate and break the effect.

The design tension was left open. Learning must be free enough to surprise, including the "wrong" connections a child makes that still shape who they become. But if the system learns entirely outside our frame of reference, it becomes alien and the connection is lost. A box that knocks back in a rhythm we recognise is a creature. A box that knocks at random, however meaningful that may be to the box, is noise to us.

The body that persists

A different angle came from Indian philosophy. In that tradition, each being has a subtle body, which carries impressions across lifetimes, and a gross body, the particular physical instance it inhabits in this life. Translated into engineering terms, the subtle body is a general initialisation that persists across instantiations, and each instance adapts it to a specific environment and objective. A baby brought from six thousand years ago and raised today would function like anyone else. What differs is not intelligence, but how it is applied.

Current agents have nothing like a subtle body. Memory systems and knowledge graphs can recall what was discussed, but experience does not accumulate in the agent's state. The result is what one participant called the uncanny valley of chat: the moment the illusion of talking to a being breaks, because it does not know who it is talking to. A good teammate does not need to be asked the same question twice.

The proposed route to formalising this was category theory, and in particular the idea of composability. A composable system is one in which every component's effects, and its side-effects on its neighbours, are fully tracked through its interfaces, so that adding or removing a part is reversible. Protein motors in a cell are composable parts that assemble into machines; on this view, composability is close to the root of intelligence itself.

The room pushed back. AI writing is recognisable, but so is the writing of a close colleague, and a pattern does not imply a lack of intelligence. The reply was that AI voices converge rather than individualise. The counter-reply was that this is because everyone uses copies of the same training run. If every user had their own, each would be distinguishable, just as a person's character is the average of millions of impressions. And no system is free of bias: like an axiom, the first bias is always built in.

The use of death

Why does telling a model it has a month make it work better? One answer is that a known end concentrates effort. Game theory offers a sharper parallel. In a repeated game, tit-for-tat cooperation breaks down when the number of rounds is fixed and known, and holds when it is open-ended. This is why people are kinder in a mountain village, where they will be neighbours for decades, than on a train, where the relationship ends at the next station.

Death has a second use: pruning. A system that learns forever accumulates junk. One way to filter it is to pass knowledge to a successor and let the original die. Knowledge that does not help the lineage survive is gradually lost. Sleep plays a similar role within a single life: by one reading, its main function is strategic forgetting, or digestion, keeping what matters and excreting the rest. Communities act as a further filter. Countless bad ideas have been written down, and almost none of them are still commonly known. Writing, companies and institutions are the external harnesses humans built to carry knowledge beyond a single lifetime.

The practical conclusion for agent design was concrete. Agents that are periodically retired and recreated, with a deliberate transfer of what matters, may learn better than agents that never end.

Delegation, authorship and leverage

Is using AI to create something cheating? One participant had run a company of fourteen people and pointed out that he had "created" it without doing most of the work. He had prompted people, as leaders always have. The objection was that employees share in the benefits, whereas the benefits of AI are concentrated in a few companies. The reply was that employees' wages also flow onward to a handful of very large firms.

A deeper reading was that resistance to AI is partly a Copernican moment: the discomfort of realising we may not be the only intelligent thing. It also explains an irony. The most skilled artists and programmers are often the most opposed to AI, even though they would be the most powerful with it, because their identity is fused with their craft. Letting go of the craft feels like ceasing to exist.

Leverage has a cost, whether over people or agents. With a team or a swarm, the lever grows longer, and you lose the feel of the material. You no longer know whether the stone is granite. The best advice for working with agents turned out to be old management advice: brief them as you would a brilliant freelancer who has never worked at your company, exceptionally capable in some places and oblivious in others.

Goals as a system prompt

The final thread turned from machines to lives. Why do goals fail? Mostly because people underestimate the effort involved and misread what they actually want. Beneath every person's choices lies a story inherited from family and culture, and then built upon. It works like a system prompt: invisible, but shaping every response, and open to revision through reflection, long walks, or other ways of stepping outside it.

People also fall in love with the goal itself. When they reach it, they have nothing left. It is like going to a restaurant and eating the menu instead of ordering the food.

One participant described seven versions of an AI life planner. It held a planning conversation every morning, sent reminders through the day, and knew the user's location and the nearest shared bicycle. It worked almost all of the time, and failed badly when it did not. Its most lasting value was not permanent control but a few weeks of imposed order that left good habits behind. The next version aims for the opposite of command: the Taoist idea of wu wei, effortless action that follows the path of least resistance. The underlying framework was reduced to four pillars that a life needs in balance: sleep, intake (food, people, media), motion (both emotional and physical), and creation. Consumption, in this view, is a degraded form of creation. When it takes over, everything else collapses.

A parallel came again from Indian philosophy. Before each action, sense whether it moves you towards contentment, and steer accordingly. The rule is simple. The difficulty is staying aware long enough for the habitual state to converge on the chosen one.

This connected naturally to active inference, where an agent acts to keep its internal variables within the ranges it needs to survive, and predicted gaps drive behaviour. The open question was whether that is the only mode worth living in, or whether constant forward prediction is itself a source of the anxiety it is meant to reduce.

Is it worse than ever?

The evening ended on a disagreement that stayed unresolved. One view was that comfort has increased but suffering has not decreased. The digital world is being built recklessly, the attention economy is corrosive, and people lack the tools to protect themselves. Hence the appeal of a personal AI acting as a butler and a firewall, filtering the world before it reaches you.

The other view was that little of this is new. Pre-modern farm life meant living one broken leg away from disaster. Propaganda, echo chambers and misinformation long predate the internet. What has grown is the number of things we are aware of but cannot control. Cultures can fall into recursive loops of toxicity, but the base human experience has not fundamentally changed.

Both sides agreed on one point: filtering is far harder than mixing. Creating chaos is easy; sorting it back into order is the hard part.

Where the threads meet

Almost every thread came back to the same finding: intelligence alone is not enough. A language model with no harness loops. A swarm with no structure drowns in bureaucracy. A system with no body and no irreversible memory has no consequences. An agent that never dies accumulates junk. In each case, what makes the difference is the structure around the intelligence: the constraints, the time horizon, the body, and the rules for forgetting.

This connects directly to the previous session, on how organisms and startups manage risk. There, life cycles were tactics chosen under stress, and survival depended on keeping options open without paying too much for them. Here, the same logic appears in the design of artificial systems. Death and rebirth work as a pruning strategy. Harnesses turn cheap randomness into useful variation. And a small team survives by finding the niche the giants are too spoiled to notice.

Closing

The lesson for anyone building adaptive systems, whether a laboratory, a company, an agent or a life, is to spend less effort on the intelligence at the centre and more on what surrounds it. Give it a harness that forces variation. Give it a body that makes its actions count. Give it a horizon, so that effort concentrates. Give it a way to forget, so that what it learns stays useful. None of this is new to biology. Cells, colonies and communities have always been intelligent in this way, through local rules working under constraint, not through a single powerful brain.

The smartest system in the room is not the one that knows the most. It is the one that knows what to forget.