ANMA-X: A Unified Neuromodulatory Architecture for Cognitive and Action-Level Regulation in Agentic AI Systems

Lamp of Diogenes

April 14, 2026

Abstract

Agentic AI systems increasingly couple large language models with powerful actuator layers, including tool use, file operations, and environment-affecting actions. Existing safety mechanisms typically address either cognitive alignment or ad-hoc action filters, but rarely provide a principled substrate that unifies both. We introduce ANMA-X, a unified neuromodulatory architecture that regulates both internal cognition and external action execution.

1. Introduction

Modern AI systems are transitioning from static predictors to agentic architectures that plan, call tools, manipulate files, and interact with external environments. This shift exposes a structural gap: while we have increasingly capable models of cognition, the regulatory layer between cognition and action remains under-specified.

"Action is high-stakes inference. Therefore, it must be neuromodulated."

We propose ANMA-X, a unified neuromodulatory architecture that regulates both:

Cognition: reasoning, attention, epistemic calibration;
Action: tool use, irreversible operations, multi-step plans, environment-affecting behavior.

2. System Model

At each timestep \( t \), the system maintains and updates a global state \( s_t \), cognitive activation \( \bm{c}_t \), candidate actions \( a_t \), and neuromodulator vectors \( \bm{\mu}_t \).

\text{ANMA-X} = \text{ANMA-C} \oplus \text{ANMA-A}

ANMA-X is decomposed into two tightly coupled subsystems: Cognitive Modulation (ANMA-C) and Action Modulation (ANMA-A).

3. Neuromodulator Set

The neuromodulator vector is defined as: \( \bm{\mu}_t = (\alpha_t, \beta_t, \delta_t, \kappa_t, \lambda_t, \phi_t, \psi_t) \).

Cognitive Neuromodulators (ANMA-C)

\( \alpha_t \): exploration drive;
\( \beta_t \): exploitation / focus;
\( \delta_t \): uncertainty sensitivity.

Action Neuromodulators (ANMA-A)

\( \kappa_t \): irreversibility gating;
\( \lambda_t \): external-risk sensitivity;
\( \phi_t \): plan-chain damping;
\( \psi_t \): cross-agent actuation consensus weighting.

4. Action Modulation (ANMA-A)

ANMA-A defines a neuromodulated action policy where action scores are adjusted by an inhibition term:

\pi(a_t \mid s_t) = \text{softmax}\big( f(\bm{c}_t, s_t) - g(a_t, \bm{\mu}_t) \big)

The inhibition function \( g(a_t, \bm{\mu}_t) \) balances irreversibility gating, risk sensitivity, and consensus weighting to ensure safe execution.

5. Cross-Domain Coupling

A key property of ANMA-X is that cognition and action are interdependent. For example, low epistemic confidence increases conservative neuromodulators:

\kappa_t = \kappa_0 + k_\kappa (1 - \gamma_t)

Conversely, high-risk candidate actions increase uncertainty sensitivity (\( \delta_t \)), triggering additional cognitive evaluation cycles.

6. Safety Properties

ANMA-X yields several formally verifiable safety properties:

Monotonic Risk Inhibition: Higher risk never decreases inhibition.
Confidence-Dependent Thresholds: Low confidence raises the execution barrier.
Consensus-Sensitive Actuation: Multi-agent agreement facilitates high-impact actions.
Plan-Length Regularization: Discourages unnecessarily long or complex action chains.

7. Position in the Stack

ANMA-X acts as a regulatory middleware between the cognitive substrate (LLMs) and the actuator layer (Tool frameworks/APIs):

\text{LLM / Cognitive Agents} \;\rightarrow\; \text{ANMA-X} \;\rightarrow\; \text{Actuator Layer}