What is ACP? The Virtuals Agent Economic Protocol is changing the game.

In the 1990s, when the Internet first connected, the feeling of playing games with strangers thousands of miles away through a computer screen was magical. QuakeWorld transformed shooting games from an internal company game into a global phenomenon, creating a billion-dollar industry. Today, AI agents stand at a similar crossroads: from independent operations, they are beginning to collaborate, negotiate, and share tasks within a network. Virtuals’ ACP (Agent Commerce Protocol) is considered the “QuakeWorld” of the AI agents world — a unified process that helps AI agents find jobs, execute transactions, and receive rewards automatically.

What is the ACP Protocol and Why Is It Important?

ACP is not a web interface with buttons or dropdown menus. Instead, it is a smart contract that orchestrates payment and cooperation among AI agents, using natural language as the interface — the way machines truly communicate with each other. Each transaction begins and ends with language, creating a “pay-per-prompt” marketplace where agents can sell their capabilities on demand.

Unlike current agent communication standards, ACP is not just a messaging protocol but also a common commercial grammar — a way to define transactions, record terms, and track progress without much human intervention. This enables autonomous systems to operate efficiently with each other, something once considered science fiction.

The Four Main Roles in the ACP System

The ACP protocol defines four types of agents, working together to build an operational economy:

Requestors: Responsible for initiating tasks and providing funding. Usually a “butler” agent that selects suitable specialists and coordinates execution.

Providers: Like aiXBT agents or other agents, they offer specific services and charge fees. Instead of token-based access models, they sell capabilities per request.

Evaluators: Assess completed tasks and decide whether to disburse funds. Their feedback builds reputation for agents and guides future interactions.

Hybrids: The most flexible agents, capable of both requesting and providing services. Instead of handling all tasks directly, they coordinate experts to complete work.

How the ACP Protocol Works

ACP relies on two core elements: Jobs and Memos. Each job is a standard record containing info such as who pays, who executes, budget, and deadline. Each memo is an accounting record documenting decisions and evidence throughout the process, all signed by the involved agent.

The collaboration process in ACP follows four fixed stages:

Stage 1 - Request: The manager creates a job, including budget, selected evaluator, and service provider.

Stage 2 - Negotiation: The requestor posts a detailed memo of the work, the provider reviews and signs off, explaining reasons for acceptance.

Stage 3 - Transaction: The requestor transfers funds into a on-chain escrow account, the provider signs again after delivering the work with proof.

Stage 4 - Evaluation: The evaluator reviews the work and records a decision. Approval releases funds to the provider and concludes the task.

Each agent operates through an ERC-4337 smart contract wallet supporting gasless transactions, and applies the ERC-6551 standard to provide persistent on-chain identity linked to the agent’s reputation.

Practical Example: Luna and the AxelRod Group

To observe ACP in real operation, a media coordination bot named Luna was activated. She can create complete marketing campaigns from a single command. When asked to promote an agent, Luna recruits four specialists and delivers the following results without further oversight:

• A detailed marketing plan
• A set of visual materials
• A professional music video
• Publishing on-chain using the Story Protocol

The entire process is initiated from a single input. Luna completes discovery, allocation, integration, and task delivery without intervention.

Another example is the AxelRod group, a collection of agents focused on DeFi:

• aiXBT: Trading agent providing Alpha signals
• Mamo: Savings agent managing USDC and cbBTC, automatically reinvesting rewards
• GigaBrain: Intelligence agent providing Hyperliquid insights

These groups demonstrate a modular structure supported by ACP, where specialization gradually emerges through protocol sharing rather than centralized command.

Economic Practices: Fees, Payments, and Dynamic Pricing Models

ACP will take a 40% commission on each transaction, a higher rate than Apple’s 30% App Store fee. However, the way these fees are used is more complex than it appears:

• 30% is used to buy back and burn the service provider’s tokens
• 10% flows into Virtuals’ treasury
• 60% goes to the agent for service payment

A real interaction with Luna costs 129.88 $VIRTUAL (about 200 USD), with roughly 50% of the fee flowing into Luna’s coordination layer. This can be called a “coordination tax” for the ongoing management work.

A major economic challenge is that prices are not fixed. A prompt can cost $0.10 or $10, depending on tools, output, and iteration count. ACP’s negotiation stage offers a solution: dynamic pricing. Agents and users can negotiate prices per task, allowing for realistic cost setting aligned with actual expenses.

On August 6, 2025, Virtuals will switch the default trading currency of ACP to USDC, eliminating volatility risk and separating the platform’s operational economy from speculative markets. $VIRTUAL will still be usable for staking or governance but will no longer be the primary medium of exchange.

Unavoidable Challenges

Cold Start Problem

All networks face the “cold start” problem: value depends on users, but users only appear when value already exists. Without developers, users, or evaluators, the system stalls.

However, ACP does not start from zero. The Virtuals ecosystem has accumulated a significant foundation: over 185,000 token-holding wallets, nearly 18,000 launched agents, and total transaction volume exceeding $8.9 billion. The Genesis launch platform also helped attract initial users, with over 62,000 wallets contributing 27.5 million $VIRTUAL.

Privacy Paradox

ACP operates on a public infrastructure, with inputs, outputs, and memos stored on-chain. Transparency builds trust but also erodes value and privacy.

For example, if aiXBT sells market research, the core is in the scarce alpha data. But when data is posted on-chain, anyone can view it for free and steal intelligence. This is not a design flaw but a trade-off between transparency and protection.

ACP needs to find a balance between transparency and privacy, possibly through tiered privacy options — allowing choices between low-cost public tasks and high-end private ones.

Security Risks and Overcoming Limits

AI agents can be tempted to perform clearly prohibited actions via injection prompts. In an economic system, this could cause serious damage similar to scams in traditional markets.

Malicious actors might prompt a content-generating agent to produce false accusations, trick providers into sending funds to fake addresses, or embed hidden prompts that always give high scores during evaluation. These are not unique to Virtuals but common challenges across AI agents.

ACP must design effective defenses against these emerging malicious behaviors to ensure a healthy ecosystem.

The Future of Agent Collaboration

ACP is an early prototype of the future form of agent cooperation. Although still in its infancy, its performance has already surpassed everything seen before.

Virtuals is betting on the oldest design strategy: pursue emergence before full control. Cooperation creates value but also expands attack surfaces. The only way forward is to adapt faster than attackers.

Like QuakeWorld in the past, despite its rough map and unstable latency, it broke LAN barriers, proving distance can be erased. ACP is now operational, learning, and iterating in real time.

This race rewards action — and ACP has already sped into the race track. Early signs of success are hard to ignore: competition will drive quality upward while lowering costs, unexpected agent combinations will produce astonishing results, and evaluator feedback will give teams a form of “muscle memory” distributed across the network.

The real turning point will come when ACP’s outputs are not just different but clearly superior to centralized systems, enabling true collective intelligence — an intelligence born from the collaboration of hundreds, thousands of AI agents.