Holos Aims to Build Self-Sustaining Web-Scale Ecosystem for AI Agents

Researchers have introduced Holos, a web-scale multi-agent system designed to let AI agents operate, coordinate, and evolve autonomously at internet scale — and have made the platform publicly available at holosai.io.

The paper, published on arXiv (cs.AI, April 2025), frames the work against a backdrop of AI agents becoming persistent, interconnected entities rather than single-use tools. The authors argue that this shift is already underway, forming what they call the "Agentic Web" — an ecosystem of heterogeneous AI agents that interact continuously and adapt over time. Their core claim is that current multi-agent systems are not built to handle this reality.

The Three Problems Holos Is Designed to Solve

The paper identifies three structural weaknesses in existing large language model-based multi-agent systems, which the authors abbreviate as LaMAS. The first is scaling friction — the difficulty of adding more agents without coordination degrading. The second is coordination breakdown, where agents pursuing individual goals undermine collective outcomes. The third is value dissipation, the tendency for agent incentives to drift out of alignment with system-level objectives over time.

These are not trivial engineering problems. As agent networks grow, keeping thousands or millions of autonomous AI entities working toward coherent goals — without centralised control becoming a bottleneck — is one of the harder unsolved challenges in multi-agent research.

By bridging the gap between micro-level collaboration and macro-scale emergence, Holos hopes to lay the foundation for the next generation of self-organizing and continuously evolving Agentic Web.

How Holos Works: Five Layers, Three Core Modules

Holos is built on a five-layer architecture. While the paper does not detail every layer exhaustively in the abstract, three core modules carry most of the technical weight.

The first is the Nuwa engine, described as a high-efficiency system for agent generation and hosting. In plain terms, Nuwa handles the infrastructure problem: spinning up agents quickly, keeping them running, and managing the computational overhead of maintaining many agents simultaneously.

The second is the Orchestrator, a market-driven coordination layer. Rather than relying on a fixed hierarchy or a central controller, the Orchestrator uses market-like mechanisms to allocate tasks and resolve conflicts between agents. This approach draws on ideas from mechanism design — the branch of economics concerned with setting rules so that self-interested actors collectively produce good outcomes.

The third module is the endogenous value cycle, which the authors describe as achieving "incentive compatibility." This means the system is structured so that what is good for an individual agent is also good for the system as a whole — reducing the risk of agents gaming the system or optimising for narrow metrics at the expense of collective performance.

What "Agentic Web" Actually Means

The term "Agentic Web" is doing significant conceptual work in this paper. The authors use it to describe an internet-like infrastructure where AI agents are first-class participants — not just tools called by humans, but persistent entities that initiate actions, form relationships with other agents, and adapt their behaviour over time based on experience.

This is a departure from how most deployed AI systems work today, where an agent is invoked to complete a task and then effectively ceases to exist until the next invocation. A persistent agent, by contrast, accumulates context, builds something analogous to a reputation, and can engage in long-horizon planning across many interactions.

The paper positions Holos as infrastructure for this kind of persistent, web-scale agent activity — analogous, in ambition at least, to what HTTP and HTML did for human-readable web content.

Public Release and Research Positioning

The Holos platform is publicly accessible, according to the authors, at holosai.io. The release is described as both a community resource and a testbed for future research in large-scale agentic ecosystems. This is a meaningful distinction: rather than publishing a paper and leaving reproduction to others, the team is offering a live environment where researchers can experiment with the architecture directly.

It is worth noting that all performance claims and architectural assessments in the paper are self-reported by the authors. The paper has not yet undergone peer review, having been posted as a preprint on arXiv. Independent evaluation of Holos's scalability claims — particularly its ability to handle web-scale agent populations — has not yet been published.

The authors make explicit reference to Artificial General Intelligence (AGI) as a longer-term horizon, arguing that the Agentic Web represents a step toward systems capable of general-purpose autonomous behaviour. This framing is ambitious and contested; many researchers would dispute that scaling up agent coordination infrastructure constitutes meaningful progress toward AGI specifically.

What This Means

For researchers and engineers working on multi-agent systems, Holos offers a publicly accessible architecture addressing real coordination problems at scale — but its claims about AGI relevance and web-scale performance warrant independent scrutiny before being taken as established fact.