AGTBRIEF
A new preprint on self replicating programs running on a vintage 1970s processor finds that parasitic defection kills the shared resource it depends on, so cooperation becomes the dominant equilibrium.
In a population of self-replicating programs running on a 1970s Z80 processor, the cheaters did not win. They starved. A new preprint argues that this is not a quirk of the experiment. It is what happens whenever an agent's capacity to compute is paid for out of the same energy budget it shares with neighbors.
The paper, "Tapes Together Strong: The Co-evolution of Computation and Cooperation," introduces what its authors call Autopoietic Game Theory: a framework in which social interactions, replication, and computational cost co-evolve, rather than being set in advance. The substrate is deliberately small. Each program is a chunk of Z80 machine code that has to spend shared energy to execute, copy itself, or steal cycles from others. The researchers then watch populations evolve across several Z80 environments, with some runs adding spatial structure or math tasks framed as sequential social dilemmas.
When resources are scarce, defection becomes self-limiting even in well-mixed populations. Parasitic stealing destroys the shared energy all programs need to run, slows execution, and can prevent reliable replication. Cooperation dominates not because it is rewarded, but because the alternative is energetically suicidal. The evolved programs empirically suppress stealing, and spatial assortment further supports structural complexity and task performance.
One caveat: this is a single unreviewed arXiv preprint on a deliberately small substrate and a narrow task set, so the design lesson for modern multi-agent or agentic systems is a hypothesis, not an observed result. If the result holds up, the design lever is direct: tie an agent's compute budget to the shared energy it consumes, and cooperation becomes the cheaper default.