A newly published editorial in the International Journal of Networked and Distributed Computing argues that the field of distributed intelligence is undergoing a fundamental structural transition, evolving beyond the Internet of AI Agents toward what the author defines as a Society of Agents. Written by Patrizio Dazzi, the manifesto observes that historical computing paradigms expanded connectivity through the Internet of Things, embedded localized decision-making through the Artificial Intelligence of Things, and subsequently envisioned decentralized environments through the Internet of AI Agents. However, the author asserts that sheer network connectivity is no longer sufficient to govern large-scale, heterogeneous populations of autonomous software entities operating on behalf of individuals, institutions, and industrial machinery. As artificial intelligence agents become increasingly capable of planning, delegating, verifying, and executing consequential actions across distributed systems, the primary engineering obstacle shifts from establishing basic communication channels to governing collective behavior across autonomous computational collectives.
To formalize this organizational shift, the editorial presents a foundational definition of a Society of Agents as a networked population of autonomous artificial intelligence entities whose collective actions are shaped by shared protocols, differentiated roles, persistent memory, operational norms, incentives, trust mechanisms, causal accountability, and governance frameworks. The author breaks down this definition across five fundamental dimensions: the agents are networked across digital channels, maintain operational autonomy rather than executing rigid static workflows, occupy structured and non-interchangeable roles, maintain persistent state across time, and remain governed by verifiable policies and sanctions. Crucially, the manifesto emphasizes that this social framing is structural rather than anthropomorphic. Artificial agents require neither subjective consciousness, biological emotions, nor human social identities to form a functioning society. Instead, the author looks to human institutions as historical proof that collective intelligence scales through organizational coordination, division of labor, shared archives, contractual agreements, and accountability systems rather than isolated individual cognitive growth.
The conceptual core of the manifesto is organized around six guiding theses designed to reorient distributed computing research toward governable collective intelligence. The first thesis dictates that the primary unit of scientific analysis must transition from individual agent autonomy to population-level collective agency, evaluating systemic interaction patterns rather than isolated capabilities. The second thesis asserts that connectivity must yield to explicit governance, establishing formal rules regarding who possesses authority to act under designated policies and supervisory controls. The third thesis emphasizes that ephemeral message exchanges must be replaced by shared societal memory capable of preserving historical commitments, decisions, and errors. The fourth thesis states that simple execution logging is inadequate, requiring robust causal accountability to reconstruct how actions produce specific outcomes. The fifth thesis calls for the evolution of workflow orchestrators into computational institutions with rules, sanctions, and dispute mechanisms, while the sixth thesis warns that individual agent safety does not ensure collective safety, demanding population-level safeguards against systemic cascading failures.
Addressing the foundational requirements of this transition, the editorial identifies shared societal memory as the first core capability necessary for sustained multi-agent collaboration. Contemporary artificial intelligence systems typically rely on isolated context windows, localized scratchpads, private logs, or fragmented vector databases, which severely limit longitudinal cooperation. The author notes that while private retrieval and planning mechanisms support basic agentic behavior, true social intelligence requires governable, persistent, and provenance-aware shared memory architectures. Drawing on established organizational memory theory, the editorial delineates four distinct memory categories required by agent societies: episodic operational memory tracking specific tasks, plans, and outcomes; semantic collective memory structuring shared domain ontologies and reusable abstractions; institutional memory encoding operational norms, policies, and procedural rules; and reputational memory recording historical reliability, compliance, and quality metrics across cooperating entities to foster decentralized cooperation.
While shared societal memory transforms fleeting interactions into cumulative collaboration, the editorial cautions that communal memory stores also introduce substantial vulnerability and expand the systemic attack surface. Unregulated shared memory can easily become stale, inconsistent, poisoned by malicious actors, or overtrusted by downstream automated components. If erroneous assertions, outdated facts, or adversarial injections enter the shared repository, false assumptions can quickly cascade through interacting agents and become institutionalized across the broader computational society. To counter these vulnerabilities, the manifesto advocates for the implementation of a formal memory constitution. This computational framework establishes explicit operational rules governing how memories are generated, validated, indexed, accessed, revised, and permanently deprecated. The memory constitution must enforce strict access controls, verify evidentiary provenance, manage temporal validity, resolve contradictory assertions, and regulate the specific operational conditions under which retained memories are permitted to direct subsequent agent activities.
The second foundational capability identified in the manifesto is causal governance, which provides the temporal ordering, attribution, and accountability structures necessary to make shared continuity trustworthy. Building on long-standing distributed systems principles such as logical clocks, vector clocks, lineage models, distributed tracing, and transactional compensation workflows, causal governance focuses on tracing how actions propagate across interacting networks. The editorial stresses that agent societies must be capable of diagnosing not merely what computational events occurred, but precisely what causal dependencies triggered them. The author illustrates this operational necessity using a representative execution sequence: when an initial agent records an assertion in shared memory, a second agent alters its tactical plan based on that assertion, a third agent executes an external tool call according to the altered plan, and a fourth agent assesses the final result to adjust peer reputations.
Without robust causal governance mechanisms in place, distributed systems can record only disconnected event timestamps, leaving developers unable to definitively determine which memory entries altered which decisions or which tool outputs triggered downstream failures. With causal governance, the entire operational dependency chain remains fully inspectable, enabling transparent fault attribution, regulatory auditing, and targeted institutional learning. To realize these capabilities, the author proposes the adoption of machine-readable causal contracts that define at design time which operational dependencies must be logged, which autonomous choices require justification, and which distributed execution chains must remain reversible or subject to direct human intervention. Such traceability is particularly critical during concurrent agent operations, where parallel task execution inevitably introduces subtle race conditions, conflicting plans, inconsistent memory modifications, and non-deterministic outcomes that severely complicate attempts to distinguish between innocent correlation and genuine computational causation.
To guide practical engineering efforts, the editorial outlines an architectural framework conceptualized as a multi-tier institutional stack situated above traditional networked infrastructure. Lower architectural layers manage essential communication, addressing, protocol negotiation, and semantic data exchange, while successive upper layers supply persistent memory services, causal lineage tracking, reputation assessment, and policy enforcement engines. The author emphasizes that a Society of Agents cannot be generated merely by increasing the raw count of autonomous agents active within a server cluster; rather, it requires adding persistent organizational structures that regulate interactions developmentally. This perspective aligns with classic research in electronic institutions and agent-oriented software engineering, illustrating how explicit computational constraints transform uncoordinated, potentially erratic multi-agent systems into stable, inspectable environments capable of maintaining predictable behavior even when individual software agents exhibit unpredictable autonomous reasoning.
The editorial highlights that the necessity for societal governance is already becoming evident in emerging multi-agent orchestration software, citing Steve Yegge’s open-source Gas Town workspace manager as an informative operational signal. Designed to coordinate multiple coding agents working concurrently across complex software repositories, Gas Town assigns differentiated responsibilities, including supervisory and subordinate worker roles, while persisting state across discrete tasks. Although the author characterizes such orchestration tools as early, domain-specific engineering experiments rather than complete academic benchmarks, their practical implementation immediately surfaces institutional governance challenges: determining work attribution, isolating which agent introduced a bug, routing tasks based on reliability, and supervising delegated authority. The author notes that this same evolution will inevitably emerge across other demanding industrial sectors, including distributed robotics, medical systems, industrial automation, cybersecurity, smart cities, and edge computing environments, whenever parallel agents modify shared resources.
In its concluding analysis, the manifesto details an extensive research agenda for the networked and distributed computing community, encompassing scalable capability discovery, robust memory governance, causal observability, multi-faceted trust models, machine-readable normative languages, collective safety protocols, and hybrid human-agent oversight frameworks. Because collective agent populations can spontaneously generate shared conventions, systemic feedback loops, and cascading failures that do not manifest in single-agent testing, safety must be investigated at the population level. The author stresses that agent societies will not operate in complete isolation from human organizations, requiring clear escalation pathways, transparent explainability interfaces, and rigorous accountability boundaries. The editorial concludes that while the Internet of AI Agents provided the essential infrastructure for digital connection, the Society of Agents represents the crucial organizational frontier required to engineer durable, trustworthy, and governable networked intelligence.
Subject of Research: Governance frameworks and shared architectures for autonomous multi-agent artificial intelligence systems
Article Title: From the Internet of AI Agents to the Society of Agents: A Manifesto for Governed Networked Intelligence
Article References: Dazzi, P. (2026). From the Internet of AI Agents to the Society of Agents: A Manifesto for Governed Networked Intelligence. International Journal of Networked and Distributed Computing, 14(1), Article 19. https://doi.org/10.1007/s44227-026-00107-1
Image Credits: AI Generated
DOI: 10.1007/s44227-026-00107-1
Keywords: Society of Agents, Internet of AI Agents, autonomous agents, causal governance, shared societal memory, multi-agent systems, distributed computing, causal accountability, episodic operational memory, semantic collective memory, institutional memory, reputational memory, memory constitution, causal contracts, multi-agent orchestration, computational institutions, population-level safety, networked intelligence
News Source: Denise Maddox. (November 19, 2026). Editorial Calls for Governed Computational Societies to Coordinate Autonomous AI Agents. Scienmag.



