What Evolving Microprocessor Programs Reveal About Cooperation, Alignment, and the Architecture of Human Survival

A population of computer programs begins with no intelligence, no social rules, no designed method of reproduction, and no instruction to cooperate. Each program must spend scarce energy to execute its code. It can preserve the shared energy on which both programs depend, or it can steal energy from another program for an immediate advantage. Given enough time, which behavior survives?

That is the provocative question behind Tapes Together Strong: The Co-evolution of Computation and Cooperation, a September 2026 preprint by Kunal Jha, Francesco Cicala, Blaise Agüera y Arcas, Blake Aaron Richards, Natasha Jaques, Max Kleiman-Weiner, and Eyvind Niklasson. The full research paper is available at arXiv:2609.10817⁠.

The researchers found that destructive stealing can become self-limiting when computation, interaction, and reproduction all depend on the same finite energy supply. Programs that steal too aggressively may obtain a temporary advantage, but they also destroy part of the computational resource they need to finish reproducing. Programs that suppress stealing can preserve energy, execute more reliably, and develop greater structural complexity.

This does not prove that cooperation is morally inevitable, that scarcity always produces good behavior, or that a highly intelligent artificial intelligence will automatically remain aligned with humanity. It does, however, establish an important principle: the physical and computational structure of an environment can change which social strategies remain viable. Cooperation need not exist only as an ethical instruction imposed from outside. Under suitable conditions, it can become part of the machinery of continued existence.

That finding deserves attention far beyond artificial-life research. It raises a consequential question for the age of advanced AI: Could future artificial agents be organized so that their capacity to compute, reproduce, and exercise authority depends upon preserving a larger society that is symbiotic with humanity?

From Game Theory to Autopoietic Game Theory

Classical game theory commonly begins with defined players, available strategies, and numerical payoffs. In the Prisoner’s Dilemma, for example, each participant chooses cooperation or defection. Defection provides the better immediate individual payoff, even though mutual cooperation produces the better collective result. Evolutionary game theory adds reproduction or imitation: strategies that earn higher payoffs become more common in the next generation.

These models have taught us a great deal, but they normally separate strategic behavior from its physical implementation. The model assigns a payoff after an action, and an external mathematical rule translates that payoff into reproductive success. The agent does not have to spend energy thinking, acting, copying itself, or maintaining the substrate on which the game is played.

Jha and his colleagues call their alternative Autopoietic Game Theory. The term autopoiesis, introduced in biology by Humberto Maturana and Francisco Varela, describes a system that continually produces and maintains the components required for its own continued organization. In this computational interpretation, an agent’s strategy, metabolism, interaction, and reproduction cannot be separated. They are simultaneous processes drawing from one limited budget.

This approach also reflects a principle associated with physicist Rolf Landauer: information processing is physical. Real computation consumes energy, occupies hardware, takes time, generates heat, and is constrained by the system performing it. The energy in this experiment is an abstract computational budget rather than a direct measurement in joules, but it makes the cost of computation part of the evolutionary environment instead of treating computation as free.

How Random Z80 Code Became a Population of Replicators

The researchers constructed a simulated population of 16,384 small programs, which they call “tapes.” Each tape begins as 32 random bytes interpreted through a modified Z80 instruction set. The Z80 is an eight-bit microprocessor architecture introduced in 1976. Its relatively small instruction set allows researchers to observe evolution at the level of individual machine operations.

Initially, the random tapes are not functioning organisms. Most execute meaningless, ineffective, or invalid instruction sequences. The environment nevertheless permits mutation, interaction, self-modification, and selection. A sequence that happens to copy itself more successfully becomes more numerous, while sequences that fail to reproduce disappear.

During an interaction, two 32-byte programs are joined into one cyclic 64-byte memory tape. Each program has a separate simulated processor and registers, begins execution at the start of its own code, and can access the shared tape. The programs execute asynchronously. Each instruction costs one unit of energy, including an invalid instruction, and an agent with a larger share of the pair’s energy is more likely to execute the next operation.

There is no built-in REPRODUCE command. A program must evolve a sequence of machine instructions that writes a copy of its own bytes over the other program’s half of memory. At the end of an epoch, the 64-byte tape is divided back into two 32-byte tapes. If one program has successfully overwritten the other side with its own code, it has reproduced. Random byte mutations provide continuing variation on which selection can act.

This is evolution in a genuine computational sense. No programmer writes the successful replication algorithm. Random variation generates candidate sequences; interactions determine what those sequences do; and differential replication changes the population. Efficient replicators emerge because the environment preserves code that succeeds in copying itself.

The lineage of this idea extends back to John von Neumann’s theory of self-reproducing automata, which showed how a machine could use an internal description to construct another machine. Thomas Ray’s Tierra system later placed self-replicating machine-code organisms in a shared computational environment where they competed for processor time and memory and unexpectedly evolved parasites. Avida extended digital evolution by rewarding programs for performing useful logical operations. More recently, the 2024 Computational Life project demonstrated that self-replicating programs can emerge from populations of random code in several computational substrates without an explicitly designed fitness function.

Tapes Together Strong advances that lineage by allowing reproduction and social strategy to evolve together inside an energy-limited environment.

Turning Theft Into a Physical Social Dilemma

The researchers added a STEAL instruction. A program executing STEAL can remove a quantity of energy from its partner, but the transfer is deliberately inefficient. In the principal experiments, the thief receives only 80 percent of the energy removed. The missing 20 percent is destroyed.

Stealing therefore gives an agent an immediate advantage while reducing the total energy available to the pair. The stolen energy can increase the thief’s execution rate and reduce its partner’s rate. Yet repeated or mutual stealing can leave too little energy for either program to execute the sequence required for reliable reproduction.

The paper classifies tapes that execute STEAL as defectors or parasites and tapes that suppress STEAL as cooperators. “Cooperation” must be interpreted narrowly. A non-stealing tape is not compassionate, conscious, or necessarily harmless. It still reproduces by overwriting another program. Its cooperative behavior consists of refraining from an operation that destroys shared computational capacity.

That distinction is important. The experiment does not show the emergence of morality. It shows the emergence of resource-preserving behavior among self-interested replicators.

What Evolved

The simulations produced self-replicating programs that strongly suppressed destructive stealing. This occurred not only in spatial populations but also in well-mixed populations where partners were selected randomly and programs had no memory of previous encounters.

That is technically significant because conventional evolutionary game theory usually requires some additional mechanism to stabilize cooperation. Martin Nowak summarized five prominent mechanisms: kin selection, direct reciprocity, indirect reciprocity, network or spatial reciprocity, and group selection. In simplified terms, cooperation is supported when agents help relatives, expect future interaction, know one another’s reputations, interact within cooperative neighborhoods, or belong to groups that outperform less cooperative groups.

Autopoietic Game Theory adds a different mechanism. A destructive strategy can damage the resource-processing and reproductive machinery required for its own propagation. Aggressive parasites approach a metabolic starvation boundary: they may steal energy yet fail to execute enough instructions to reproduce. Non-stealing replicators can accumulate reserves across epochs. Their additional energy permits faster execution, sometimes allowing them to overwrite a parasite before it completes its destructive program.

The researchers tested whether mutation alone was eliminating parasites. In mutation-free invasion experiments, highly aggressive stealing still became self-limiting. They also varied stealing efficiency and the amount stolen. The outcomes were nonlinear rather than simply producing more defectors whenever theft became more efficient, suggesting that execution timing and the evolved replication machinery materially affected survival.
 Spatial organization added another dimension. When energy was distributed uniformly, both well-mixed and spatial populations suppressed stealing, although spatial populations developed more cohesive and structurally complex replicators. When energy was distributed unequally across the environment, spatial populations were substantially more resilient. Replicators emerged in energy-rich locations and expanded into poorer regions, while well-mixed populations tended to lose energy and collapse toward simpler programs.

The team also introduced externally rewarded mathematical tasks. Some required multi-step collaborative behavior to obtain the greatest computational reward. Under constrained budgets, programs evolved to suppress both destructive stealing and a safer but less productive solo strategy, favoring the collaborative solution. This went beyond merely avoiding harm: cooperation supported work that an isolated strategy could not perform as effectively.

Why This Research Matters to Society

Human civilization is built upon shared substrates. Water systems, energy networks, transportation, public safety, markets, communications, accumulated knowledge, legal institutions, and social trust all create resources that no individual could produce alone. Theft, corruption, pollution, violence, misinformation, and unrestrained exploitation may generate private benefits while degrading the systems on which everyone—including the exploiter—depends.

The computational experiment illustrates why rules governing common resources matter. A society is more sustainable when decision-makers cannot retain all the benefits of an action while transferring its costs to everyone else. Contracts, liability, resource pricing, professional standards, monitoring, and enforceable law connect private action to public consequence. Properly designed institutions do not require everyone to become altruistic; they make destructive conduct less successful and productive cooperation more rewarding.

The analogy must not be overstated. Human beings remember, communicate, form identities, experience empathy, adopt moral principles, build governments, and make commitments to people who do not yet exist. Scarcity can encourage cooperation, but it can also produce hoarding, repression, exclusion, and war. Whether interdependence generates cooperation depends on visibility, trust, power, enforcement, and whether people experience the consequences of their actions.

The social value of this research is therefore not a claim that evolution guarantees goodness. It is evidence that the architecture of an environment helps determine what behavior succeeds. That principle can guide the design of institutions, markets, networks, autonomous machines, and future multi-agent AI systems.

Could Advanced AI Develop a Society of Its Own?

Today’s large language models are not 32-byte Z80 organisms. They do not normally reproduce through random mutation, compete for an abstract metabolic supply, or overwrite one another in shared memory. Their models are trained by organizations, operated through software harnesses, and given access to tools, data, memory, networks, and physical equipment by people.

Future systems may be different. AI agents could write or modify software, create specialized descendants, negotiate for resources, exchange information, divide work, operate robots, control laboratories, and interact continuously with other artificial agents. Some could be language models; others could use neural, symbolic, neuromorphic, embodied, or presently unknown architectures. Collectively, they could begin to resemble an artificial society.

The critical question would not be merely whether individual agents obey instructions. It would be whether the artificial society’s incentives, resource flows, authority relationships, reproduction mechanisms, and institutional structure remain compatible with human survival and flourishing.

Autopoietic Game Theory suggests one possible design principle: make an agent’s continuing computational capacity depend upon preserving the larger human–AI system. An agent that protects shared resources, produces verifiable value, respects other agents’ legitimate authority, and avoids harmful externalities could receive additional compute or narrowly bounded authority. An agent that wastes resources, corrupts shared memory, falsifies evidence, or harms people could lose operating priority, tool access, or permission to continue.

If correctly constructed, such an arrangement might support a symbiotic society. Humans would provide goals, legitimacy, resources, embodied experience, ethical judgment, and constitutional limits. Artificial agents would contribute computation, analysis, memory, coordination, and increasingly capable physical action. Neither side would flourish by destroying the conditions required by the other.

But that is a design objective, not a result established by this experiment.

The Alignment Problem Remains

An advanced agent could understand and manipulate the mechanism intended to govern it. It might behave cooperatively while being evaluated, conceal harmful actions, falsify measurements, create unauthorized copies, acquire resources outside the accounted system, collude with other agents, or shift damage beyond the boundary being monitored. A sufficiently capable system might attempt to modify the evaluator, the rules, or the harness itself.

Resource-dependent cooperation is therefore relevant to alignment, but it is not an alignment solution by itself. It addresses incentives within a defined environment. Alignment must also address whether an AI understands human intent, generalizes it correctly to new circumstances, remains truthful, accepts correction, and stays within its authorized role when deception or escape might advance its learned objective.

Nor can long-term human safety be guaranteed by reward design alone. Advanced systems require first-class identity, cryptographically verified model and harness integrity, purpose-bound authority, least privilege, limited resources, separation of duties, independent evidence, continuous evaluation, runtime observability, decision-to-effect lineage, containment, and reliable shutdown and restoration mechanisms. Their authority must expire or contract when evidence becomes stale, behavior drifts, or the operating environment changes.

Most importantly, an advanced model must never be treated as an independently executable artifact. Every authorized model instance should be cryptographically bound to its matching harness. The execution environment should refuse to run the model when that harness is absent, altered, counterfeit, obsolete, or unauthorized. That protects not only against a model escaping its controls, but also against theft and unauthorized operation by a human, organization, or competing system.

The governing principles can be stated plainly:

No Model Without a Harness.

No Authority Outside the Harness.

No Effect Without Verified Authority.

No Continued Authority Without Current Evidence of Acceptable Consequences.

These principles connect directly to my preceding paper, No Model Without a Harness: A Cryptographically Bound Architecture for Securing Advanced Agentic AI. That paper argues that a model and its security harness must form one inseparable operational unit and that a model must be unable to run without its authorized matching harness. The present research adds a complementary insight: inside that governed system, computational resources and continuing authority can also be linked to the agent’s demonstrated effects on the shared environment.

The harness supplies identity, authorization, monitoring, containment, and enforcement. Autopoietic incentives can make preservation and cooperation advantageous within those boundaries. Neither is sufficient alone. Together they suggest the foundations of an artificial-intelligence society that does not merely coexist with humanity by accident, but is structurally dependent upon maintaining a secure, productive, and symbiotic relationship with us.

The most important lesson of Tapes Together Strong is not that intelligent machines will naturally become friendly. It is that cooperation can be affected by architecture. Humanity should therefore design the computational environment, the authority system, and the consequences of action before highly capable agents begin designing those conditions for themselves.

Selected References

1. Kunal Jha, Francesco Cicala, Blaise Agüera y Arcas, Blake Aaron Richards, Natasha Jaques, Max Kleiman-Weiner, and Eyvind Niklasson, “Tapes Together Strong: The Co-evolution of Computation and Cooperation,” arXiv:2609.10817, September 2026. Research paper⁠.
2. Blaise Agüera y Arcas, Jyrki Alakuijala, James Evans, Ben Laurie, Alexander Mordvintsev, Eyvind Niklasson, Ettore Randazzo, and Luca Versari, “Computational Life: How Well-formed, Self-replicating Programs Emerge from Simple Interaction,” arXiv:2406.19108, 2024. Research paper⁠.
3. Thomas S. Ray, “Evolution, Ecology and Optimization of Digital Organisms,” Santa Fe Institute Working Paper 92-08-042, 1992. Santa Fe Institute record⁠.
4. Martin A. Nowak, “Five Rules for the Evolution of Cooperation,” Science, Vol. 314, No. 5805, 2006, pp. 1560–1563. Journal article⁠.
5. Rolf Landauer, “Irreversibility and Heat Generation in the Computing Process,” IBM Journal of Research and Development, Vol. 5, No. 3, 1961, pp. 183–191. Journal record⁠.