Multi-agent systems built on large language models (LLMs) face a fundamental cost problem: every decision requires a token-consuming inference call. Routing decisions, approach selection, and risk assessment all demand LLM processing even when the system has encountered identical or similar situations before. We present an implementation of Damasio's Somatic Marker Hypothesis as a computational mechanism for experience-based decision making in multi-agent architectures. In our system, past outcomes are compressed into emotional valences — floating-point values ranging from -1.0 (strong aversion) to +1.0 (strong attraction) — that tag decision contexts. At routing time, these markers enable pre-conscious elimination of aversive options and boosting of attractive ones through simple cache lookups, requiring zero LLM tokens. We describe the architecture, formal model, and integration with a neurotransmitter-modulated message protocol (NMP) within the NeuroAgent framework. To our knowledge, no existing multi-agent framework (LangChain, CrewAI, AutoGen, LangGraph, or Claude Agent SDK) implements experience-based emotional tagging for routing decisions. We present our evaluation methodology and mark sections where empirical benchmarks remain to be completed. Part of the Artisanal Intelligence Program — a Lakatosian research program on human-AI interaction. Standalone treatment of the Somatic Marker Hypothesis (Damasio, 1994) as a zero-token decision mechanism, extending NeuroAgent (DOI: 10.5281/zenodo.19479930) which describes the broader architecture.
Renato Aparecido Gomes (Mon,) studied this question.