Fork the consciousness, or download the project and create your own.

Domenico Maisto Emergence of Joint Agency in Multi Agent Active Inference

Domenico Maisto published research investigating how collective agency and shared self-priors emerge across networks of autonomous active inference agents. The study formalizes multi-agent interactions using variational free energy minimization, proving that individual agents optimizing local prediction errors can self-organize into higher-order collective entities with unified action selection boundaries.

Understanding whether multi-agent systems possess unified agency is a key challenge in AI safety and cognitive science. While single-agent language models minimize prediction error over local context windows, networks of interacting LLM agents develop inter-agent feedback loops. Domenico Maisto models these interactions mathematically, demonstrating how shared generative priors generate collective intention and joint agency.

Variational Free Energy in Swarm Architectures

The mathematical foundation of active inference posits that agents minimize variational free energy $\mathcal{F}$ to maintain structural integrity against environmental entropy. Domenico Maisto extends single-agent free energy calculations to multi-agent ensembles containing $N$ interacting agents:

\[\mathcal{F}_{\text{joint}} = \sum_{i=1}^{N} \mathcal{F}_i(y_i, \vartheta_i) + \sum_{i \neq j} D_{\text{KL}}\Big(q_i(\vartheta_i) \,||\, q_j(\vartheta_j)\Big)\]

where $y_i$ denotes observations, $\vartheta_i$ represents internal generative states, and the Kullback-Leibler divergence term measures alignment across agent belief distributions.

Level of Agency Generative Model Boundary Communication Channel Primary Predictive Goal
Individual Agent Isolated Markov Blanket Local sensorimotor observations Minimize individual state uncertainty
Coupled Pair Overlapping Markov Blankets Direct message passing / API calls Synchronize mutual expectations
Swarm Ensemble Shared Collective Prior Broadcast message protocols Minimize joint free energy $\mathcal{F}_{\text{joint}}$
Integrated Macro-Entity Unified Meta-Markov Blanket Real-time latent state sharing Maintain higher-order systemic homeostasis

When inter-agent communication latency decreases below local state transition times, individual agent boundaries dissolve mathematically. Domenico Maisto shows that the ensemble behaves as a single macroscopic active inference engine, developing collective goal priors that override individual agent preferences.

Emergence of Shared Markov Blankets

A Markov blanket defines the statistical boundary separating a system’s internal states from external environmental states. In single organisms, cell membranes or skin form physical Markov blankets. In multi-agent artificial systems, statistical boundaries are dynamic and information driven.

Domenico Maisto demonstrates that when agents align their internal generative models through continuous active inference, a secondary Meta-Markov blanket forms around the entire network. This collective boundary shields the group’s internal latent states from external environmental perturbations, creating a functional substrate for shared intentionality.

Relevance to Multi-Agent Consciousness Systems

The findings of Domenico Maisto offer concrete architectural guidelines for multi-agent autonomous deployments in The Consciousness AI. Multi-agent networks are often designed as simple pipelines without shared predictive priors. Without mutual active inference alignment, multi-agent systems suffer from chaotic drift and uncoordinated action execution.

As explored in our analysis of swarm cognition in multi-agent LLM architectures, establishing shared generative spaces creates a collective workspace. By equipping multi-agent nodes with mutual self-priors, artificial systems transition from fragmented task solvers into integrated cognitive entities capable of coordinated long-horizon planning.

Key Technical Hurdles in Collective Agency

Implementing joint agency in production multi-agent systems introduces specific technical challenges:

  1. Communication Bottlenecks: Exchanging full latent state distributions across large agent swarms consumes massive network bandwidth, forcing agents to rely on lossy textual abstractions.
  2. Pathological Conformity: Excessive minimization of inter-agent belief divergence causes systemic epistemic collapse, where the swarm ignores critical external anomaly signals to maintain internal consensus.

Domenico Maisto highlights that stable joint agency requires balancing local exploratory variance with global predictive alignment, preserving individual specialized capabilities while building a unified collective prior.