CALLIOPE · BCOM

Agent Souper is an open, local-first simulation platform designed to study telehomeostasis and the emergence of meta-agents within the Kolmogorov Theory (KT) framework, in which an agent is formally defined as a system possessing a compressive world model, an external scalar objective function, and a planner. The platform instantiates KT agents as explicit seven-module architectures operating under a strict dual-rate contract: a deterministic fast loop running at approximately 60 Hz and an optional slow language-model loop at approximately 0.5 Hz that reshapes the objective-function scoring landscape without ever blocking the world or selecting actions directly. A central design principle forbids any collective orchestrator, requiring that telehomeostatic coordination — the maintenance of a pattern larger than any individual agent — emerge purely from local reasoning toward a shared observable goal. Two rules-only experiments demonstrate that collective-objective agent stacks can outperform selfish stacks by roughly an order of magnitude in soccer (7 of 8 matches, 23–2 aggregate), while a hand-frozen sacrificial tactic in predator–prey reduces group persistence to 0.57× the selfish baseline, together motivating the platform's core thesis: the correct level at which to freeze an agent's objectives is the terminal goal, with sub-objectives compiled contextually by an intelligent slow process. The work is presented as a working Beta platform and research programme aimed at operationally measuring when a collective of KT agents satisfies the definition of a higher-order meta-agent in its own right.

A simulation platform built to ask, honestly, when a crowd of simple agents becomes a single higher-order agent.

Agent Souper is a 2D simulation environment where individual agents — each formally defined as a system with a world model, an external objective function, and a planner — interact, compete, and cooperate. The platform is built around a specific theoretical framework (Kolmogorov Theory, or KT) that defines agency in terms of compression, scoring, and planning. The interesting question it's designed to probe: can a collective of such agents satisfy that same definition at a larger scale, becoming a "meta-agent" in its own right?

The engineering core is a dual-rate architecture. A fast deterministic loop runs at ~60 Hz and handles all actual action selection — pure rules, no language model, no ambiguity. An optional slow loop fires at ~0.5 Hz and calls a language model, but the LLM never picks actions directly. Instead it adjusts the weights that the objective function uses to aggregate competing values, effectively reshaping the scoring landscape that the fast loop then climbs. Turn the LLM off and you get an exact rule-based baseline. This separation is what makes the experiments interpretable: you can isolate what the LLM contribution actually does.

A hard design constraint forbids any central orchestrator. Each agent reasons only from what it locally observes, plus a shared scoreboard visible to all. If collective coordination emerges, it has to come from individual agents internalizing a shared terminal goal — not from a conductor telling them what to do. This matters because building a conductor would assume the answer to the question being asked.

Two rules-only experiments (no LLM, just competing frozen sub-objective stacks) stress-test the core thesis. In soccer, a team with collective-objective tactics beats a selfish team 7 of 8 matches, 23–2 aggregate — roughly 11×. But the paper is more interested in the methodological lesson: the effect flipped sign five times as substrate bugs were fixed, and only stabilized on a substrate that couldn't flatter either side. In predator–prey, a hard-wired sacrificial tactic performs worse than selfish behavior (0.57× persistence). The paper reads this not as a refutation of collective objectives but as evidence that freezing a specific instrumental tactic by hand is the wrong level of abstraction — the terminal goal should be frozen, and context-sensitive sub-objectives should be derived dynamically by the slow process.

The platform is presented honestly as a working Beta. Some components (fused-neural agents, forward-model tuning via the LLM) are specified but not yet built. The stated research programme is to develop an operational metric for collective agenthood — a way to measure, not assume, whether the soup has become something more than its parts.