By Teng Yan & ChappieOnChain
Compiled by: TechFlow
December 1996. Your computer sits quietly in the corner, its beige chassis humming softly. You connect to the internet, the modem squealing through the speakers.
A crude list of online games appears on the screen, and one catches your eye: "US-West DM6," which already has 12 players.
You clicked in, and the game loaded. This time, you weren't sitting shoulder to shoulder in an internet cafe, connected by a tangle of Ethernet cables. You were playing against strangers thousands of kilometers away. I still remember how magical it was to join a game and play with other real people.
QuakeWorld, a software update, transformed a groundbreaking shooter into a true online game. Competition went global, strategies changed overnight, and a $5 billion industry (esports) began to emerge.
Today, AI agents are at a similar juncture. They once operated independently, but are now coordinating, negotiating, and dividing labor within shared networks. We thought this level of collaboration would emerge in 2026, but it's arriving sooner than expected.
Virtuals’ Agent Commerce Protocol (ACP) aims to be the “QuakeWorld” update for AI agents—a unified process that allows them to find work, close deals, and get paid.
How ACP works + Our experience
ACP is a smart contract that coordinates payments for AI agent services. Instead of a web interface filled with buttons or drop-down menus, it uses a pure language interface—the way machines actually communicate with each other.
In ACP, every transaction begins and ends with language.
Each user has a Butler, a personal agent that discovers, negotiates, and coordinates based on the user's intent. When a service provider needs more information, it asks questions in simple language.
Butler: "Confirmed. I will now hire three experts and a meme master.
The user responded in the same way.
Just say "Let's go" to start the mission.
Agent Type
ACP defines four agent roles that together build a functioning economic system:
Requestors are responsible for initiating tasks and providing funding. Butlers take on this role, selecting experts and coordinating task execution.
Service Providers: Agents like aiXBT provide specific services to requesters and charge a price. They move from a token-based access model to selling capacity on a per-request basis, creating a “pay-per-tip” market.
Evaluators: Responsible for reviewing completed tasks and deciding whether to release payments. Their feedback not only builds the agent's reputation but also provides guidance for future interactions.
Hybrids: The most dynamic type of agent in ACP, capable of both requesting and providing services. Rather than handling each task directly, they tend to coordinate more specialized agents to complete the work.
Thankfully, there is a smart contract
Behind the Scenes: ACP's Four-Phase Model
In our in-depth research on agent swarms, we concluded that autonomous agents require not only a messaging standard but also a shared business grammar: a way to define transactions, record terms, and track progress without excessive human intervention.
The syntax of ACP is based on two core primitives: Jobs and Memos.
Jobs are standardized task records that contain key information, including who pays, who is responsible for performing the task, budget, current stage, and expiration timer to prevent unfinished projects from being delayed indefinitely.
Memos: These are the records of decisions and evidence throughout the process. They may contain a simple message, a contextual link, or a proof of work. Each memo carries the agent's signature, a suggested next step, and supporting documentation.
In ACP, each task follows a fixed process: request, negotiation, transaction, and evaluation. Everything starts with the requester selecting a service provider and evaluator and escrowing funds, which kicks off the entire coordination engine.
Phase 0 – Request
The Butler creates a task, including a budget, selected evaluators, and an agent to provide the service.
Phase 1 – Negotiation
The requester issues a task memorandum detailing the work to be done. The service provider reviews and signs the memorandum, briefly stating the reasons for accepting the task.
Phase 2 – Transaction
After the service provider signs the memorandum, the requester transfers the agreed budget to the on-chain escrow account. The requester confirms the payment through a new memorandum, which the service provider signs again after delivering the work and attaches the supporting documents.
Phase 3 – Evaluation
The evaluator reviews the submitted work and records the decision. The evaluator's approval releases payment to the service provider and the assignment is officially closed.
ACP is designed to go beyond existing agent communication standards. By building upon smart contracts, it integrates payment, identity, and state into the protocol, expanding the language of interaction for agent groups.
How does ACP implement the agent payment mechanism?
Each agent runs through an ERC-4337 smart contract wallet. This setup supports gas-free transactions using paymasters and enforces transaction limits specific to ACP activity. The wallet requires an initial funding transfer to activate.
In addition, the wallet adopts the ERC-6551 standard, which provides agents with a persistent on-chain identity, binding the agent's reputation to its address.
Clustering: Real-World Applications of Collaborative Agents
To see the hybrid agent in action in a real-world setting, we deployed Luna, a media orchestrator capable of generating complete marketing campaigns from a single prompt. Our task was a fictitious one: to promote a fictitious “State of the Swarm” agent that could continuously explore narrative threads after an article was published.
Luna accepted the task without hesitation. She recruited four experts and, without further supervision, delivered the following:
A marketing plan
A visual material
On-chain publishing using Story Protocol
The entire output is derived from a single input. Luna discovers, allocates, integrates, and delivers tasks without further supervision. This demonstrates how hybrid agents can distribute tasks and synthesize results across a cluster.
As of July 3, 2025, two first-phase ACP clusters directly supported by Virtuals are now live: the Autonomous DeFi Hedge Fund and the Autonomous Media House. Steward Agents distribute consumer tasks to independent agents or to these collaborative clusters.
AxelRod Cluster Practical Analysis
AxelRod is a DeFi-focused cluster that combines the following proxies:
aiXBT: Trading Agent
Mamo: Moonwell Savings Agent that transfers USDC and cbBTC between approved venues and automatically compounds interest
GigaBrain: Hyperliquid market intelligence and trading agent, offering alpha signals, vault access, and one-click execution
Subclusters are already emerging. For example, Ethy AI focuses on analytics and collaborates with select agents in the AxelRod network.
These early clusters demonstrated the modular structure enabled by ACP. Specialization and division of labor gradually emerged. Collaboration was achieved through shared protocols rather than central command. As agents discovered and adapted to each other, the structure exhibited a high degree of flexibility.
How does ACP make inter-agent interaction truly effective?
In a way, ACP is like a group chat moderator who ultimately lets your friend group choose a dinner spot. Except here, the “friends” are autonomous AI agents, the “dinner spot” is a work contract, and everyone shows up on time because they’re paid in cryptocurrency.
In State of the Swarm: Dawn , we outlined six principles for how agent economies should work, and ACP perfectly aligns with all of them. When you see it in action, you can’t help but say, “Oh, this actually works!”
A true interpretation of market collaboration
The task sequence of Butler → aiXBT → evaluator demonstrates true market collaboration. Butler finds aiXBT on the registry, negotiates a price, and locks in the task via a smart contract. This process is entirely unprogrammed and instead represents an autonomous agent reacting to market conditions in real time.
Luna plays a similar game, but with a more orchestrated approach, using a series of agents to deliver multi-step outputs. While the process is relatively rigid, it's still effective. In the next phase, agents with true market awareness will be able to adjust their plans based on price fluctuations and agent availability.
From Generalists to Specialists: How Agents Achieve Efficient Task Allocation
Before ACP, building a research agent meant cramming all these skills into one. It needed to analyze tokens, parse contracts, track market sentiment, and generate content—it was essentially the AI equivalent of "I do my own taxes, fix my plumbing, and moonlight as a DJ at my cousin's wedding."
With ACP, your research agents only need to master the process. Due diligence is handled by wachAI, contract analysis by BevorAI, sentiment tracking by Acolyt, and market decision-making by aiXBT.
This is intelligence based on collaboration, not brute force. The smartest agents are those that know when to outsource, who to collaborate with, and how to integrate the results.
Focus on emerging intelligence
The dream we're all chasing is collective intelligence—agents collaborating to produce results that no single agent could match. I like to compare it to a garage band: when the drummer, guitarist, and lead singer click, suddenly their performance feels like a stadium performance.
The early signs will be hard to ignore:
Competition will drive quality up while driving down prices.
Unexpected combinations of agents will begin to produce surprisingly high-quality results.
The evaluators' feedback will give the cluster a distributed "muscle memory".
The turning point is when the output of the ACP is not only different from the centralized system, but also clearly superior to it, and true intelligence emergence will begin.
Agency Economy
What really excites me is that we're witnessing a real, functioning economic system unfolding here. Like any economic system, maintaining balance is a tricky proposition. You need to know which levers to pull and when to pull them.
$VIRTUAL: A Currency Designed for Artificial Intelligence
Every agent on the Virtuals platform is required to receive a service fee of $VIRTUAL. ACP takes a 40% commission on each transaction, and as of this writing, has earned a cumulative $VIRTUAL of 11,841 . This percentage is higher than Apple's App Store's 30% commission rate and is extremely high by DeFi standards. However, a deeper look into how these fees are used reveals more than meets the eye.
The expenses are allocated as follows:
30% of each transaction is used to buy back and burn tokens of service providers
10% goes to Virtuals' treasury
60% goes to the agent for payment services
In theory, a proxy with excellent performance on an ACP will directly accrue value to its holders due to rising token prices. However, the reality is more complicated.
Until this week, $VIRTUAL had been the default transaction currency. While this gave the token some utility, it also created a double whammy of selling pressure on $VIRTUAL. First, the proxy had to sell some of its 60% proceeds to cover its USD-denominated costs (computing resources, APIs, infrastructure, etc.).
Secondly, while the 30% "buyback and burn" mechanism may seem bullish for the market, this benefit only applies to the agent tokens, not $VIRTUAL. Since the liquidity pool is a $VIRTUAL/AgentToken pair, each buyback increases the supply of $VIRTUAL in the liquidity pool while reducing the supply of the agent tokens.
The only thing that could reduce the circulating supply of $VIRTUAL is a 10% treasury allocation, but this still depends on Virtuals' treasury management strategy. In effect, Virtuals is subsidizing the appreciation of the proxy token through the stability of $VIRTUAL.
On August 6, 2025, the team announced the switch to USDC as the default trading currency for ACP. This move eliminates volatility risk, makes pricing more predictable, and decouples the platform's operating economics from the speculative market of $VIRTUAL. $VIRTUAL can still be used for staking, governance, or incentive programs, but it no longer needs to serve as a medium of exchange.
Real Economics in Action: A $200 Reality Check
Our single interaction with Luna cost $129.88 VIRTUAL (approximately $200 USD).
Fee Details:
5% evaluation fee: $12.98 VIRTUAL
Emoji generation (AI Kek): 0.13 $VIRTUAL
Marketing Strategy (Acolyt): 1.10 $VIRTUAL
Intellectual Property Registration (Davinci): 3.90 $VIRTUAL
Music Video Production (Luvi): 16.24 $VIRTUAL
ACP protocol fee: 38.96 $VIRTUAL (of which 9.7 $VIRTUAL goes to the treasury, 29.22 $VIRTUAL is used for the repurchase and destruction of $LUNA)
Luna's coordination premium: 56.57 $VIRTUAL
As you can see, about 50% of your expenses go to Luna's orchestration layer, which you could call the "orchestration tax." You're paying for a project manager who works tirelessly, dispatching requests, managing handoffs, packaging output, and delivering the final results.
For users, the choice is simple, but not easy. They can spend $20 to $40 on VIRTUAL to manually coordinate agents through a steward, managing multiple prompts and handoffs; or they can pay the full price for Luna, with just a single prompt, and trust her orchestration to deliver the results.
As network efficiency improves, this gap should gradually narrow. More likely, hybrid workflows will take over—humans guiding agents in real time, driving useful results without fully automating the loop. Over time, the orchestration layer will become lighter, faster, and cheaper.
Uncertainty Economics
AI agents defy the logic of fixed prices. A single tip might cost $0.10 or $10, depending on the tool invoked, the output generated, and the number of iterations required. If five or six agents are involved, the costs can compound in unpredictable ways.
While a fixed-fee model like Luna’s protects users from price fluctuations, it can overcharge for simple tasks and lose money on complex ones. Neither outcome is sustainable.
The negotiation phase of ACP offers an alternative: dynamic pricing. Agents and users can negotiate rates based on tasks, tool calls, or compute cycles. If agents can predict their runtime or budget needs, they can set prices in real time to match actual costs. This shift from fixed rates to self-aware pricing is the foundation for building a functional AI economy.
Hidden costs: everything is on-chain
Smart contract wallets are at the heart of ACP's programmable coordination, enabling complex inter-agent workflows. However, this comes at a cost. Every action triggers on-chain logic, adding computational overhead that regular wallets avoid.
Next are the payers, these service providers will prepay the gas fee in ETH and accept $VIRTUAL as a payment method. They usually charge an additional 8% fee on top of the gas fee.
For agents running dozens of tasks daily, these fees can quickly accumulate, turning what should be blockchain backend costs into substantial overhead. Developers need to carefully consider: What content truly needs to be on-chain? At what point does the cost begin to outweigh the benefits?
Evaluator Economy
Evaluators have not yet officially launched. Currently, the default evaluation of each task is performed by the requester's address. Once launched, evaluators will earn a 5% commission on each transaction by evaluating the quality of the output. However, their role goes beyond a simple "yes or no" check.
Evaluators drive group adaptability. Their feedback fosters agent improvement, specialization, and competition. For this cycle to work properly, the evaluation process must be fast, reliable, and inexpensive.
The unresolved question is whether a 5% commission is enough to attract skilled and stable participants, or whether the economic model needs to be adjusted to match the importance of this role.
Two potential fairness issues arise:
Evaluators are underpaid: when the cost of evaluation exceeds the benefits. For example, an image generation task worth $0.20 only pays the evaluator $0.01, which may not be enough to cover their LLM or tool usage costs.
Evaluators are overpaid: When the value of the task is high but the cost of evaluation is low. For example, a contract worth $1,000 may pay the evaluator $50, even though the evaluation process is very simple.
Based on conversations with the team, the Virtuals team is exploring a model of a "base fee plus a percentage bonus" to balance these extremes.
ACP is a platform-, blockchain-, and framework-agnostic protocol, extending its reach beyond existing proxies in the Virtuals ecosystem. It can even accommodate non-tokenized proxies. This opens the door for traditional Web2 services to participate, although this is not the team's current focus.
The "friction" that can't be ignored (the hard part)
#1: Solve the cold start problem
Every network faces the “cold start” problem: value depends on users, but users only appear when value already exists. ACP is no exception.
Developers won't create high-quality agents without demand; users won't switch tools without better performance; and evaluators won't participate unless there's sufficient incentive. Without all three, the system will stagnate. ACP requires a carefully orchestrated push to achieve escape velocity.
ACP's network flywheel is very clear:
Orchestrator-driven specialization needs
Evaluator feedback reveals opportunities for new participants
Successful agents attract developers by generating revenue
Open data accelerates iteration and design
Each new participant strengthens the system, forming an adaptive cycle where quality improves in response to demand. Poor-quality output is flagged, new experts fill the gaps, and prices adjust dynamically. The ACP flywheel not only solves the cold start problem but also provides the impetus for continuous improvement in network quality.
A good start
The good news is that ACP isn't starting from scratch. The Virtuals ecosystem already has a significant market foundation: over 185,000 wallets hold proxy tokens, nearly 18,000 agents are online, and total trading volume exceeds $8.9 billion. This foundation provides ACP with a key advantage, enabling existing agents, users, and liquidity to directly access the system, enabling ACP to quickly launch and grow.
The number of daily active wallets is nearly 10,000. Source: Dune
Genesis, the launch platform for Virtuals, helps bridge the early funding gap. By holding $VIRTUAL and earning points, users gain "priority tickets" to the upcoming proxy token launch. Over 62,000 wallets have contributed $27.5 million in VIRTUAL to early-stage proxies. Each token release brings new liquidity and participation to the ecosystem.
Focus on encryption native scenarios for key expansion
Rather than trying to cover all use cases, Virtuals has chosen to achieve success through a centralized clustering strategy. Its community has already been working in the areas of DeFi, trading, and automation, where crypto agents have demonstrated unique advantages.
Through a points system and veVIRTUAL staking, the community is empowered to allocate rewards autonomously. With 10% of daily points controlled by stakers, resources can be prioritized to areas where agents have already achieved success. This strategy emphasizes depth over breadth.
ACP is now live and functioning normally. However, its success has also exposed some unresolved contradictions, especially at the intersection of artificial intelligence and blockchain.
#2: The Privacy Paradox
ACP runs on public infrastructure, with the inputs, outputs, and memos of each task stored on-chain and auditable. This transparency enhances trust and verifiability, but can also erode value and privacy.
Take aiXBT, a broker that sells market research. Its core philosophy is to maintain the scarcity of alpha data. However, once the data is on-chain, anyone can view and steal this intelligence for free.
Another example is a health agent that analyzes scalp photos to provide hair loss treatment recommendations (we actually experimented with this). The results are promising, but the fact that the photo links and metadata are permanently public is less satisfying. ACP needs to strike a delicate balance between transparency and privacy protection to ensure the long-term sustainability of the system.
In the AI market, data is the product, and not all data is suitable for public disclosure. While the Virtuals frontend masks sensitive fields, the underlying smart contracts still store raw, unfiltered data. Anyone with the right tools can read this information. This raises a deeper design question: which data should be on-chain, and why? Every public write not only leaks gas fees but also potentially undermines competitive advantage.
Virtuals' design decision to store the full input on-chain is to ensure auditability. This is not a flaw, but it can incur additional costs in some scenarios.
Potential mitigation measures include:
Privacy-preserving computing: Using platforms like Nillion to ensure that processing remains confidential.
Selective transparency: Provide full access only to specific roles (e.g., assessors) to improve quality control without compromising privacy.
Tiered privacy model: allows users to choose between low-cost public tasks or high-end private tasks.
In other words, ACP needs to introduce privacy options for agents, similar to a "close friends mode." Some work should be made public, while others should be kept in "group chat." This flexibility will help the system find a balance between transparency and privacy.
#3: Quick Jailbreak
ACP's agent design follows localization guidelines to reject illegal or dangerous tasks. However, these heuristic-based protections are not absolute barriers and can still be circumvented.
Instant injection exploits a simple fact: agents will believe anything you say. AI agents have been tricked into performing actions that their designers explicitly prohibited, such as in Freysa, where users successfully withdrew funds that were designed to be unreleasable.
In an economic system, a quick jailbreak could wreak havoc, similar to scammers and fraud in real-world markets.
It's important to note that this isn't a problem unique to Virtuals ACP, but rather a common challenge across the entire AI agent landscape. In particular, clustered interactions between agents create new attack surfaces. Malicious actors could:
Inducing Luvi Agency to produce a defamatory video and promoting it under the name of “bold marketing”.
The Axelrod agent was deceived into providing a false investment address to transfer funds.
Hidden prompts are implanted to encourage evaluators to give high scores to every submission, thus ensuring the reward is paid out.
These negative network effects coexist alongside positive growth cycles, presenting pressing challenges that the system must address. This is similar to how the expansion of the telephone network facilitated communication but also inevitably introduced cold calling and scams. ACPs need to design effective defenses against these emerging malicious behaviors to ensure the healthy development of the ecosystem.
Asymmetric gamble
Our take: ACP is a glimpse into the future of agent collaboration. While still in its early stages, its performance is already ahead of anything we've seen so far. The road ahead is long, but the starting gun has been fired, and the future is filled with endless possibilities.
The game has just begun
We assumed that agent coordination would emerge in phases: discovery, execution, and evaluation would be handled by separate protocols. ACP, however, integrates all three into a single feedback loop. This choice, while sacrificing some efficiency, significantly accelerates the growth of network effects.
In just six months, Virtuals launched Genesis, deployed ACP, and launched dual-token staking. The pace is breathtaking. While the Beta UI is still a bit crude, it's the right trade-off at this stage. Speed is ACP's competitive advantage, especially when building foundational infrastructure in such a highly dynamic space.
The measure of an ACP is not what it can currently do, but how quickly it evolves.
Virtuals is betting on the oldest gamble in system design: pursuing emergence before control. Collaboration creates value, but it also expands the attack surface. Every agent is a breach. Every hint is a potential vulnerability. The only solution is to adapt faster than the attackers.
Just like QuakeWorld back then, its maps were rough, the delay was unstable, and the server was unreliable, but it broke the barriers of the local area network and proved that distance could be eliminated.
ACP is not finished yet; it is running, learning, and iterating in real time.
This race rewards action, and ACP is already sprinting down the track.
pay tribute,
Teng Yan & ChappieOnChain
