AGENT SOUPER

A simulation platform for studying whether groups of AI agents can become something more than the sum of their parts — and why freezing the wrong goal is the root of collective failure.

Agent Souper is a software platform for running populations of AI agents that interact in a shared 2D world. Each agent is built according to Kolmogorov Theory (KT), which treats cognition as: build a compressed model of the world, define a scalar objective, pick actions that maximize it. The platform's job is to make that formalism runnable so you can actually test ideas about collective behavior rather than just theorize about them.

The key architectural commitment is a strict two-speed design. A fast loop runs at ~60 Hz, is fully deterministic, and never touches a language model — it just scores candidate actions through a weighted objective function and picks the best one. A slow loop runs at ~0.5 Hz and can call an LLM, but only to re-weight the objective function's priorities or nudge planner parameters. The LLM never blocks the fast loop, never names actions directly, and can't rewrite the rules. This separation is what makes experiments interpretable: you can turn the LLM off entirely and get bit-identical, reproducible behavior from the rules alone.

The central scientific question is about telehomeostasis — the idea that a collective can persist not because some central controller coordinates it, but because each individual agent has internalized a goal defined at the group level ("the tribe survives"). The platform operationalizes this with no orchestrator at all: each agent just has a persona and objective stack oriented toward collective success, plus access to a shared observable (like a scoreboard). Two experiments test this. In 11-vs-11 soccer, a telehomeostatic team (pass forward, hold shape, press together) beats an otherwise identical selfish team 7 out of 8 seeds, 23–2 in aggregate — but only after the authors fixed substrate bugs that were artificially inflating the result, which they report as a methodological lesson in its own right. In predator-prey, a hand-coded "sacrifice yourself to draw fire" strategy for prey persists at only 0.57× the selfish baseline and loses all 8 seeds.

The predator-prey failure is the more interesting result. The paper argues it doesn't refute the idea of collective-oriented objectives — it refutes freezing a specific tactic by hand. Whether to sacrifice is context-dependent; hard-wiring it is the wrong level of abstraction to freeze. This leads to the paper's unifying argument: an agent's objective is always frozen somewhere, and the design choice is only where. Rules-only freezes the instrumental sub-goals (the specific tactics). The LLM-on regime moves the freeze up to the terminal goal ("tribe persists") and lets the slow process re-derive tactics from observation each time. The LLM experiments validating this design are described but the full empirical test of the meta-agent claim — whether a collective can itself satisfy the KT definition of an agent at a higher scale — is explicitly flagged as an open measurement program, not yet a result.