The Pro-Action Operator: The Feasibility of Bio-Inspired Regulatory Harnesses for LLM Agents

Patricio Julián Gerpe

Keywords: LLM agents, self-regulation, regulatory harness, bio-inspired AI, endogenous activation, prompt engineering, interoception, iterated prisoner's dilemma, homeostatic control, agent orchestration

TL;DR: We propose the Pro-Action operator, a bio-inspired regulatory harness that injects verbalized internal state into LLM prompts and shows feasible opponent-differentiated behavior in a 920-cell IPD benchmark.

Abstract: Autonomous LLM agents are increasingly deployed in high-stakes settings where reliability depends not only on what they can do, but on how they decide when to act, wait, defer, or escalate. Existing architectures usually delegate this activation problem to external orchestrators, rather than to an endogenous regulatory process. We introduce the \textbf{Pro-Action operator} $\Gamma$, a six-subsystem coupled thermostat used as a \textit{regulatory harness} for LLM policy execution, together with \textit{Regulatory State Verbalized Interoception} (RSVI), which exposes regulatory state to the prompt without directly prescribing actions. In a 920-cell iterated prisoner's dilemma benchmark across three providers, Full-$\Gamma$ exhibits opponent-differentiated cooperation, especially against Grim versus reciprocal opponents. These results support the feasibility of coupling LLM policy execution to explicit regulatory state, while architectural necessity, lexical priming, payoff improvement, and generalization beyond this task remain matters for further studies with matched controls.

Submission Category: Full Paper

Overaged Verification: Yes

Latin American Hispanic Heritage: Yes

Icml Proceedings Status: No

Submission Number: 23