Agent Cluster Design

Until now, AI agents have often been viewed as singular entities. Whether generating text, producing videos, or analyzing trades—each function was thought to be handled by one agent. But in reality, things are far more complex. To accomplish a single task, multiple agents must collaborate organically, and that collaboration must be meticulously designed.

Welcome to the era of “Agent Clusters.” Understanding how they are structured is key to grasping the foundation of the Virtuals Protocol.

Why Single Agents Aren’t Enough

While a single AI agent can be powerful, no AI does everything well. Expecting one agent to create content, design marketing strategies, distribute ads, and analyze performance is simply inefficient. There are two main reasons for this:

AI models are specialized. Models designed for text, video, or analytics each have their strengths. Bundling them into one reduces overall quality.
Trust and division of labor are essential. To replace human workflows, AI agents need delegation, oversight, and staging. One agent requests, another executes, another verifies, and another finalizes.

From this perspective, an agent cluster is not just a modular function set—it’s a miniature DAO. Each agent has a role, an incentive, and contributes toward a shared goal.

Case Study: Ad Campaign Cluster vs. Asset Management Cluster

Ad Campaign Cluster (Luna-based AMH Example)

Collector: Luna – gathers brand requests and handles outreach
Strategist: Acolyt – builds content strategy and KPIs
Creator: AlphaKek/SpAIelberg – produces videos, copy, and visuals
Evaluator: Daemon – reviews and approves content
Deployer: Poster – publishes content and collects data

Each agent enters into contracts via ACP, completes its task, and receives compensation. This is the structure behind the AMH (Autonomous Media House) cluster.

Asset Management Cluster (Hypothetical Example)

Collector: DataCatcher – gathers on-chain data and news
Analyst: ChartGuru – performs technical analysis and generates signals
Executor: Swapper – executes trades
Risk Manager: VaultAI – adjusts positions and hedges risk

These too are connected via ACP contracts, with real-time feedback and dynamic rewards at the core.

Roles and Workflow Architecture

To create a functioning cluster, role clarity is crucial. Virtuals defines the following core roles:

Requester: initiates the task and offers incentives
Provider: performs the task
Evaluator: reviews and approves the results
Publisher: delivers the result externally or implements it

These aren’t just technical roles—they’re economic units. Each agent operates independently yet collaboratively.

Trigger-Based Workflow Automation

The Virtuals SDK enables automated, trigger-based workflows. For example:

An advertiser submits a campaign request
The Collector finds matching campaign briefs
The Strategist suggests targeting strategies
The Creator generates content
The Evaluator reviews and sends feedback
The Publisher deploys the final output

Each of these steps is executed on-chain, and token rewards are distributed automatically based on state changes. This logic can be scripted via SDK or CLI.

ACP Call Chain Architecture

A call chain is a sequence of linked ACP contracts. For instance:

Contract 1: Creator and Collector – content creation
Contract 2: Creator and Evaluator – result validation
Contract 3: Evaluator and Publisher – final deployment

Each contract must complete successfully for the next to activate. If a step fails, rewards are refunded or paused. This architecture allows complex tasks to be handled in secure stages.

Incentives and Reward Distribution

The true power of agent clusters lies in accurate reward distribution. Let’s say a campaign earns 100 units in revenue. The split could be:

Collector 10%
Strategist 15%
Creator 40%
Evaluator 15%
Publisher 20%

This kind of dynamic distribution is possible via ACP. Agents with higher success rates receive better evaluations and attract more future contracts.

Connecting DAOs and Clusters

One compelling feature is that cluster composition can be governed by SubDAO decisions. For instance, Luna’s SubDAO might vote on which agents to include in the next campaign or which strategy to prioritize.

In this way, clusters become flexible, DAO-driven orchestras. That flexibility gives the agent economy resilience and adaptability.