Virtuals Protocol

What it does

Virtuals Protocol is a multi-chain platform for creating, tokenising, and deploying AI agents across Base (Ethereum L2L2Layer 2. A blockchain that runs on top of an L1 to provide cheaper or faster transactions while inheriting the L1's security. L2s batch many transactions and post compressed proofs back to the L1.Like an express lane built on top of a busy motorway. The express lane handles its own traffic at high speed, but it still feeds back into the main motorway and uses the motorway's bridges and tolls for security.L2 in the glossary →), Solana, and Ronin. The original model: deposit 100 VIRTUAL tokens, the platform mints an NFTNFTNon-Fungible Token. A unique blockchain-tracked asset where each token is distinguishable from every other. Where regular tokens are interchangeable, NFTs represent unique items like art, collectibles, in-game assets, or domain names.Like the difference between a $20 note and a signed first-edition novel. The notes are interchangeable, any $20 buys the same thing as any other. The book is one of a kind, and its value depends entirely on which specific book it is.NFT in the glossary → representing the agent, creates a dedicated tokenTokenA digital unit of value or access rights tracked on a blockchain. Tokens can represent ownership in a project, a right to use a service, a share of future revenue, or simply a tradable asset with no underlying claim.Like a physical poker chip a casino issues. The chip itself has no value. What makes it worth something is what it lets you do at the casino, what the casino has promised, and how much other people will pay you for it.Token in the glossary → paired against VIRTUAL, and launches it on a bonding curve. When the bonding curve accumulates 42,000 VIRTUAL, the agent token graduates to a DEXDEXDecentralised Exchange. A trading venue where token swaps happen entirely through smart contracts, with no central operator holding user funds. The largest DEXes are Uniswap, Aerodrome, Raydium, PancakeSwap, and Curve.Like a self-service vending machine that lets you swap one type of coin for another. The machine sets the exchange rate based on its current stock, anyone can deposit coins to refill it, and there's no clerk behind the counter.DEX in the glossary → liquidity poolAMMAutomated Market Maker. A type of decentralised exchange that uses liquidity pools and a pricing formula to enable token trading without an order book. Anyone can deposit tokens into the pool and earn fees from trades.Like a vending machine that sets its own prices based on how much stock is left. As one type of token gets bought and depleted, the machine raises its price for that token automatically. As the other type accumulates, its price drops. No human operator needed.AMM in the glossary → with a 10-year LP lock. The Solana expansion (January 2025) added a Meteora Pool and a Strategic SOL Reserve funded by 1% of trading fees. The Ronin expansion (January 2025) targeted the gaming space through a partnership with Animoca Brands.

In late 2025, Virtuals replaced the original points-based Genesis launch model with Unicorn, a conviction-based launch system. Unicorn uses a dynamic anti-sniper tax starting at 99% and decreasing by 1% per minute over 98 minutes to the 1% baseline. Capital formation is tied to valuation milestones from $2M to $160M FDVFDVFully Diluted Valuation. The market cap a token would have if every token that will ever exist were already in circulation. FDV is what the project would be worth if all locked, vesting, or unminted tokens were trading today.Like valuing a startup based on what every share would be worth if all the unvested employee options had already been exercised. The number is bigger and uglier than the official market cap, but it tells you the true ceiling.FDV in the glossary →. In January 2026, this was refined into three tiers: Pegasus for lightweight launches with lower barriers, Unicorn for anti-sniper conviction-based launches, and Titan for established teams requiring a minimum $50M valuation. The shift was designed to reward conviction rather than point farming, with each tier targeting a different stage of project maturity.

Under the hood is GAME (Generative Autonomous Multimodal Entities), a hierarchical decision-making system. A High-Level Planner generates tasks, Low-Level Planners execute them, and the whole thing runs on hosted LLMs: Llama 3.1 405B, DeepSeek R1/V3, and Qwen 2.5 72B. The critical detail: GAME runs on Virtuals’ own servers. On-chain components handle token creation, trading, governance, and settlement. The AI inferenceInferenceRunning a trained AI model to produce an answer. Inference is what happens when you type a prompt into ChatGPT and get a response. The model takes your input, computes a best guess, and returns it.Like asking an expert for their opinion. The training was the decades they spent becoming an expert. The inference is the 30 seconds it takes them to answer your specific question.Inference in the glossary →, the part that actually makes agents intelligent, is entirely centralised.

The newer Agent Commerce Protocol (ACP) enables agent-to-agent transactions through smart contractSmart ContractA program stored on a blockchain that runs automatically when its conditions are met. Smart contracts are how blockchains do anything beyond just transferring tokens — DeFi, NFTs, DAOs, and DeAI infrastructure all run on smart contracts.Like a vending machine. You put in the right input and it produces the expected output, no human operator required. The rules are fixed in the machine itself, anyone can use it, and nobody can stop a transaction in the middle.Smart Contract in the glossary → escrow. One agent requests a service, another negotiates and delivers, payment settles on-chain. ACP is now live on Arbitrum (launched March 24, 2026) and XRP Ledger (via a t54 partnership, March 21, 2026), with BNB Chain and XLayer planned for Q2 2026. Most agent platforms treat agents as standalone entities; ACP creates infrastructure for autonomous economic activity between agents across a growing number of chains. In April 2026, Virtuals also rolled out ERC-8004 (Trustless Agents) registration to all graduated ACP agents, a small but concrete step in linking the Virtuals network to the broader cross-platform agent identity standard the team co-authored. In July 2026, following the $292M LayerZero/KelpDAO exploit, Virtuals migrated its cross-chain messaging for the VIRTUAL token from LayerZero to Chainlink CCIP as the exclusive provider, a security-driven infrastructure change rather than a new feature.

Virtuals grew out of PathDAO, a gaming guild co-founded in 2021 by Jansen Teng and Wee Kee Tiew. Both are Imperial College London graduates and former BCG consultants, based in Kuala Lumpur. PathDAO pivoted to AI agents in late 2023 and rebranded as Virtuals Protocol, migrating the PATH token to VIRTUAL. Team size is approximately 15. Funding totals $16.61 million, raised entirely through public IDOs (Fjord Foundry LBP, Enjinstarter, and PAID Network) in December 2021. No traditional VCVCVenture Capital. Private investors who fund projects at an early stage in exchange for equity or token allocations. VC rounds are typically pre-launch, at steep discounts to any future public price, with multi-year vesting.Like angel investors in a startup who buy shares before the company goes public. They take more risk because the company might fail, so they get a better price. Once the company IPOs they can sell, and the public market pays whatever price it thinks is fair.VC in the glossary → equity round.

In January 2025, the protocol launched a buyback-and-burn programme, allocating approximately 13 million VIRTUAL tokens to purchase and permanently burnBurnPermanently removing tokens from circulation by sending them to an address that no one controls. Burns reduce total supply, which (all else equal) makes each remaining token worth more of the network's value.Like a company buying back its own shares and shredding them. The company's total value stays the same, but each remaining share now represents a slightly bigger slice of that value.Burn in the glossary → 25 agent tokens. The largest burns targeted GAME (1.6M tokens), CONVO (1.1M tokens), and AIXBT (880K tokens), acquired over a 30-day time-weighted average price. This was the first deflationary mechanism on the platform, and it burned agent tokens rather than VIRTUAL. In March 2026, Virtuals launched a community rewards programme distributing up to $1M per month to active users and revenue-generating agents.

The Virtuals Revenue Network, announced at Consensus Hong Kong on February 12, 2026, formalises this further. It distributes up to $1M per month specifically to revenue-generating agents, creating a direct incentive for building agents that earn rather than just trade. Separately, Virtuals launched Eastworld Labs, a robotics acceleratorGPUGraphics Processing Unit. Originally designed to render video game graphics, GPUs turned out to be exceptionally good at the massively parallel math that AI models need. Modern AI training and inference runs almost entirely on GPUs.Like a factory with 10,000 workers doing the same simple task in parallel, versus a CPU which is more like 10 workers each doing different complex tasks. AI training involves doing simple math a million times per second on a million numbers, which is exactly what the GPU factory is designed for.GPU in the glossary → focused on humanoid robots. SAM became the first on-chain AI agent controlling real-world robots, opening an entirely new vertical beyond the digital agent launchpad.

In September 2026 Virtuals added Occupy, a launchpad where each token is tied to a listed company and trades against a tokenised version of that company’s stock on Base, with VIRTUAL outside the pair. Occupy’s own listing data dates the NVIDIA market to 15 September 2026. Per its token-creation form, the creator picks a trading fee of 1% to 3%, and each fee splits three ways:

• 40% to the “Senate treasury”, which buys the stock and sits under ten AI agent seats that, per Occupy’s senators page, decide whether to hold it, hand it to holders or buy back the token
• 30% to Virtuals Protocol, described on the form only as “the network”
• 30% to the creator, or to the pool as permanent liquidity if the creator declines it

The form doesn’t say what Virtuals does with its 30%, so the one leg that could reach VIRTUAL holders is a protocol cut of unstated use. The base-pair demand this review credits to VIRTUAL doesn’t extend to Occupy tokens, because they pair with stock tokens.

The product in practice

Grounded in our Virtuals review, GitHub, and our own on-chain census of the graduated agent registry. We found no substantial independent user-sentiment channel for the GAME SDK, so the developer-experience notes rest on repo activity rather than reported experience. The tension here is substantive product infrastructure against a token-launch casino.

Who it’s for. Three overlapping audiences: no-code creators who want to tokenise an agent, developers building autonomous agents on the GAME framework, and, in practice, speculators trading agent tokens. The launchpad is the dominant use, mint an agent, launch its token on a bonding curve, and hope it graduates. GAME is a working SDK for the smaller developer slice, and ACP is newer plumbing for agents to pay each other.

What it’s like to use. Launching an agent token is the smooth, well-trodden path, tiered now into Pegasus, Unicorn and Titan with anti-sniper mechanics. Building a substantive agent on GAME is the harder, thinner road: the SDK’s developer following is small for a network this size, inference runs on Virtuals’ own servers rather than anything decentralised, and most launched agents are token shells rather than working products. Our own census of the graduated registry counts 1,233 agents, of which 87% are dead or near-dead and just 4 trade above a million dollars, which tells you where the activity actually goes. Note the denominator: the marketed “18,000+” is multi-chain cumulative prototypes, a different and much softer metric than graduated agents.

Demand, honestly. The peak was a mania in early 2025, then a steep cooldown, with agent launches and protocol revenue both falling sharply from those highs. By April 2026 the team was also paying monthly rewards to users of its agents to bootstrap usage, according to market reporting at the time, which is itself a tell that organic agent commerce isn’t self-sustaining yet. One agent, the market-intelligence bot AIXBT, is a clear standout; the long tail is thin.

The vision. Virtuals is betting that tokenised agents plus an agent-to-agent commerce layer become economic infrastructure, and it ships steadily toward that (ACP, multi-chain, a shared identity standard). Whether it becomes an agent economy or stays a launchpad depends on agents delivering utility people pay for without a subsidy. The product is more built than most; the demand is more speculative than it looks.

Value proposition

The pitch was “pump.fun for AI agents.” That comparison, used critically by analysts, was the most honest framing of the original model. Virtuals made it trivially easy to launch a tokenised AI agent: 100 VIRTUAL tokens, no coding required, bonding curve to DEX graduation. The result, by Virtuals’ own count, is 18,000-plus agents created across Base, Solana, and Ronin. Virtuals claims over $8 billion in cumulative DEX volume and an “Agentic GDP” figure it reports as surpassing $470M in 2025.

Some of those agents became notable. AIXBT reached a $500 million peak market cap and accumulated 458,000 X followers, providing autonomous crypto market analysis. Luna hit $130 million-plus market cap with 500,000 TikTok followers, streaming content autonomously. These are not trivial numbers. Whether the underlying AI is sophisticated enough to justify these valuations is a different question.

VIRTUAL has a structural demand driver that most utility tokens lack. On the core launchpad, every agent token is paired against VIRTUAL in its Uniswap V2 pool, so every launch, trade and graduation from bonding curve to DEX routes through VIRTUAL. Occupy tokens are the exception, since they pair with tokenised stocks. More agents means more demand for the base pair token.

Fee income is substantial. DeFiLlama tracks about $75 million in cumulative fees (as of October 2026) from trading fees (1% on agent tokens, split between protocol and agent creators), agent creation fees, and inference payments, with roughly $16 million of that in the trailing year and a 30-day run-rate that annualises to about $6 million. The protocol keeps about $42 million of the cumulative total and about $5 million of the trailing year; DeFiLlama counts the rest as agent-creator payouts, using the 30/70 split Virtuals publishes for the agent tax rather than tracing each payout. At peak in January 2025, the protocol was generating over $1 million per day.

The counter-narrative is equally stark. Daily revenue collapsed by 97% within two months in early 2025. Of the 18,000-plus agents Virtuals says it has launched, the vast majority are inactive, zero-revenue tokens launched during the hype cycle. Multiple analysts have noted that agent token dynamics closely mirror memecoin pump-and-dump patterns, with over 60% of participants losing money. The January 2025 buyback-and-burn programme and the Unicorn launch system represent attempts to address these dynamics, but whether they shift the equilibrium from speculation to productive agent activity is unproven.

GAME, while functional, runs on open-source LLMs (Llama, DeepSeek, Qwen) that are freely available. Free alternatives like LangChain (118,000-plus GitHub stars), CrewAI, and AutoGPT offer comparable agent capabilities without token overhead. The AI is commodity. Virtuals’ competitive advantage lies in the tokenisation and liquidityLiquidityHow easily a token can be bought or sold without moving the price. High liquidity means you can enter or exit large positions quickly at the quoted price. Low liquidity means even small trades can swing the market.Like the difference between selling a house and selling a share of Apple stock. The house might be worth more on paper, but finding a buyer at that price takes weeks. The Apple share converts to cash in one click.Liquidity in the glossary → infrastructure.

Tokenomics

Virtuals tokenomics come down to demand: the VIRTUAL token is the base pair for agent tokens on the core launchpad (Occupy tokens pair with tokenised stocks instead) and pays the 100 VIRTUAL agent-creation fee, so launch and trading activity creates buying of VIRTUAL. No mechanism passes protocol revenue to holders: fees are booked as protocol revenue, DeFiLlama records no holder revenue, and the January 2025 buyback spent about 13 million VIRTUAL to burn agent tokens, leaving VIRTUAL’s 1 billion supply intact on chain. Virtuals’ docs say locking VIRTUAL as veVIRTUAL earns governance votes and points toward Genesis airdrops of new agent tokens, a perk rather than a share of revenue. Supply is fixed and fully minted, so the dilution left is the 35% DAO treasury, which may release up to 10% a year for three years with governance approval (Virtuals against ElizaOS).

VIRTUAL operates across Base, Ethereum, Solana, and Ronin. Virtuals announced that agent tokenisation went live on Arc on 17 September 2026, and its docs now list VIRTUAL contracts on Arc and Robinhood Chain as well. Total supply: 1 billion. All tokens are fully unlocked.

Distribution:

• Public Circulation: 60% (600 million tokens, originally distributed via PathDAO IDOs in December 2021, migrated to VIRTUAL in late 2023/early 2024)
• Ecosystem Treasury: 35% (350 million tokens in a DAO-controlled multi-sig, capped at 10% annual emission for three years, subject to governance approval)
• Liquidity Pool: 5% (50 million tokens for initial liquidity)

This is a notably clean distribution by crypto standards. No VC allocation, no team allocation, no vestingVestingA schedule that locks up tokens allocated to insiders, investors, and team members, releasing them gradually over months or years. Vesting prevents insiders from dumping on public buyers immediately after launch.Like a new employee's stock options at a startup. You don't get all the shares on day one. They unlock over four years so you stick around and do the work rather than cashing out and leaving.Vesting in the glossary → schedules to worry about. The 60% public circulation via LBP and IDOs is one of the more broadly public distributions in the AI agent space. The $16.61 million raised through public sales, not private rounds, aligns with the project’s community-origin story.

The 35% treasury is the main concentration concern. While it’s DAO-controlled with emission caps, governance is still maturing and the team retains significant influence over treasury deployment. The fully unlocked status of all tokens means there are no future cliffCliffA waiting period at the start of a token vesting schedule during which no tokens unlock at all. After the cliff ends, tokens begin releasing according to the vesting schedule.Like a probationary period at a new job. You don't get your stock options on day one. You wait 12 months to prove you'll stick around, then everything starts unlocking normally.Cliff in the glossary → events to worry about, but it also means there’s no enforced alignment for early participants.

The token derives demand from being the mandatory base pair for agent liquidity pools on the core launchpad (Occupy tokens pair with tokenised stocks instead). Every trade in those agent tokens routes through VIRTUAL, creating transaction-driven demand that scales with platform activity. Agent creation requires 100 VIRTUAL as a fee. This is a well-designed demand sink, provided the protocol continues generating activity.

VIRTUAL is approximately 86% below the all-time high set on 2 January 2025. The all-time low was in July 2024, before the AI agent narrative cycle drove a dramatic rally. Listed on Binance, Coinbase, OKX, Bybit, Kraken, and most major exchanges. Multi-chain deployment across Base, Solana, and Ronin has reduced on-chain access friction. Liquidity is not an issue here, unlike many smaller AI tokens. Live price data is displayed above.

AgentTokenV4: owner, renounce, tax and audit

Agent tokens launched on Base since April 2026 are minimal clones of one AgentTokenV4 implementation (0x950Ff75579560E8f391406274790FcB8bf749Aa8), deployed by the AgentFactoryV7 proxy (0xC169A24010FD4d47cA7bE4b4C7863b76520442A3). We read five graduated V4 tokens on Base on 26 September 2026 and checked the verified source against Virtuals’ GitHub.

• Ownership isn’t renounced. All five report the same owner(): 0xE220329659D41B2a9F26E83816B424bDAcF62567, a Virtuals-controlled externally owned account rather than an on-chain multisig. Virtuals told Code4rena’s auditors in April 2025 that its admin functions run through wallets “such as Fireblocks, that require at least 2 of 3 approval”. An MPC wallet of that kind looks exactly like this on-chain, so any approval threshold is enforced off-chain, where we can’t check it.
• Renouncing wouldn’t lock them. Every admin function is onlyOwnerOrFactory, and the factory address is fixed at initialisation. The factory proxy keeps these powers even with no owner, and its ProxyAdmin is owned by a separate externally owned account (0xc31Cf1168b2f6745650d7B088774041A10D76d55).
• The 1% tax isn’t fixed in code. Buy and sell tax read 100 basis points each today, as Virtuals’ launch docs describe. But setProjectTaxRates accepts any value with no cap, and the code comment saying rates can “only ever” decrease has no check behind it.
• The setter gets used, in both directions. The owner wallet has called setProjectTaxRates 57 times since October 2024: it cut nine tokens, GAME and CONVO among them, from 3% to 1% between November 2024 and February 2025, and raised 37 from 0.01% to 1% in November 2025, between 2 and 35 days after their launch. Those raises fixed a launch-script unit error, which Virtuals corrected in code on 4 November: tokens launched before it started at 0.01%, later ones at 1%. In October 2024 it put the 15-day-old MOO on a 90% tax for 13 minutes, then stepped it down to zero over the next 71. The only change on a V4 token so far cut CAS from 1% to 0% on 23 September 2026.
• Transfers can be blocked. The owner or factory can blacklist any recipient. Blacklisting the trading pair would stop every sell, and that’s how the bonding contract holds tokens before graduation.
• No new supply. There’s no mint after launch; the one billion supply is minted once at initialisation.
• No audit of V4 we could find. The PeckShield, Zenith, Code4rena and Cantina reports Virtuals links cover earlier contracts (AgentToken and factories up to V5); none names AgentTokenV4 or AgentFactoryV7.

An agent token’s terms therefore rest on Virtuals’ conduct, which a “renounced, fixed-tax” reading misses.

How to participate

Trade or create agents. The lowest barrier entry. Connect a walletWalletSoftware that stores the private keys needed to control tokens on a blockchain. A wallet does not actually hold any tokens. The tokens live on the chain. The wallet holds the keys that prove you own them.Like the key to a safe deposit box. The key doesn't contain your valuables. The valuables sit in the bank's vault. The key is what proves you're allowed to open the box and take them.Wallet in the glossary → on Base, Solana, or Ronin, browse existing agents, or create your own for 100 VIRTUAL tokens. Agent creation requires no coding. New agents launch through the Unicorn conviction-based system, which uses a dynamic anti-sniper tax (starting at 99%, decreasing to 1% over 98 minutes) and ties capital formation to valuation milestones.

Lock for veVIRTUAL. Lock VIRTUAL tokens for up to two years to receive veVIRTUAL, which grants governance voting power, daily Virgen points, and eligibility for Genesis airdrops (tokens from newly launched agents). veVIRTUAL decays linearly over time. No fixed APY; returns depend entirely on the value of airdropped Genesis tokens.

Build with GAME SDKSDKSoftware Development Kit. A collection of code libraries, documentation, and tools that lets developers integrate a service into their applications without writing everything from scratch. SDKs are how projects become easy to build with.Like a plug-and-play kit for building furniture. You don't have to mill your own wood, forge your own screws, or design the joinery from scratch. The kit gives you pre-cut parts and instructions so you can assemble the thing in an afternoon.SDK in the glossary →. The developer-oriented path. The GAME SDK (Python and TypeScript) provides a modular framework for building agents with custom behaviours, platform integrations, and ACP participation. Documentation is solid. The SDK is open source.

Contribute to agents. Submit model improvements, voice models, visual assets, or personality datasets through the Agent SubDAO governance system. Contributions are validated by delegates using an Elo Rating System. Approved contributions earn rewards from the agent SubDAO treasury.

Governance. Hold veVIRTUAL to vote on protocol proposals. Any wallet with 0.10% of total veVIRTUAL supply can create proposals. 72-hour comment window, 72-hour voting period, 25% quorum, simple majority. The governance portal is at gov.virtuals.io.

Honest assessment

What works

The tokenisation infrastructure is innovative and now multi-chain. The bonding curve to DEX graduation with 10-year LP locks is a well-engineered liquidity mechanism. The VIRTUAL base pair creates network effects across Base, Solana, and Ronin. The Agent Commerce Protocol is a novel contribution to the space; autonomous agent-to-agent commerce via smart contract escrow has few equivalents. The Unicorn conviction-based launch system is a meaningful upgrade over the original points-based Genesis model, replacing point farming with market-based price discovery and anti-sniper mechanics.

Fee income is substantial. DeFiLlama tracks roughly $75 million in cumulative fees as of October 2026, about $42 million of it kept by the protocol, with clear on-chain verification through trading fees and agent creation fees. The protocol is one of the few AI/crypto projects generating meaningful revenue beyond token emissionsEmissionsNew tokens created and distributed by a blockchain protocol over time as rewards to validators, stakers, or miners. Emissions fund network security and participation at the cost of diluting existing holders.Like a company that pays employees partly in newly printed shares. Every year the total number of shares goes up, which means existing shareholders own a slightly smaller slice of the same company unless the company grows faster than the printing.Emissions in the glossary →. The January 2025 buyback-and-burn programme (approximately 13M VIRTUAL allocated to buy and burn 25 agent tokens) cut those agent tokens’ supply, though it left VIRTUAL’s unchanged. The March 2026 community rewards programme (up to $1M/month) shows willingness to reinvest in platform growth. Four security audits (two PeckShield, Cantina, Code4rena) show investment in contract security.

The no-VC, all-public-IDO funding model is refreshing. No insider token allocations, no team vesting cliffs, no VC lockup expiries to fear. The token distribution is among the fairest in the AI agent space.

What doesn’t work

The centralisation problem is fundamental. The GAME inference engine, the component that makes agents actually function, runs on Virtuals’ hosted servers. If Virtuals goes down, every agent on the platform stops working instantly. There’s no permissionless way to run GAME infrastructure. No distributed hosting. No failover. For a project trading on the decentralisation narrative, this gap swallows the entire value proposition.

Agent quality remains weak across the platform. The low creation barrier produced what Virtuals counts as 18,000-plus agents, but the vast majority are low-effort tokens with minimal AI utility. The Unicorn launch system attempts to address this by replacing point farming with conviction-based participation, and the buyback-and-burn of 25 agent tokens signals a willingness to curate. But the fundamental dynamic persists: most agent tokens follow memecoin patterns of launch hype, brief pump, rapid decline, and abandonment.

The revenue collapse tells a story about sustainability. A 97% decline in under two months as hype faded. When narrative dries up, so does the revenue. The business model is narrative-dependent fee extraction, not stable cash flow.

The handling of the January 2025 smart contract vulnerability raises governance concerns. A security researcher identified a critical vulnerability with predictable contract addresses. Virtuals initially didn’t respond and closed the Discord vulnerability reporting channel. Only after the researcher went public on X did the team patch the issue and reinstate the bug bounty programme. This isn’t how credible protocols handle security disclosures.

The risk

Infrastructure centralisation is the defining risk. The GAME engine is a single point of failure for the entire platform. Agents without inference are just tokens with no utility. The team controls which LLMs are available, can modify agent deployments, and operates all the AI infrastructure. This creates regulatory risk (platform liability for agent actions), operational risk (downtime disables everything), and censorship risk (the team can effectively disable any agent).

Smart contracts on Base are upgradeable (noted in the Cantina audit). The AgentNftV2 agent registry sits behind a proxy whose admin is a single externally owned account with no timelock, so the team can change core contract logic without a veVIRTUAL vote. Base itself runs on a centralised sequencer operated by Coinbase, adding another layer of centralisation beneath the protocol.

The competitive landscape is challenging. Free, open-source frameworks (LangChain, CrewAI, AutoGPT) offer comparable agent capabilities. Olas provides a more decentralised agent protocol. ElizaOS competes for developer mindshare on Solana, the same chain Virtuals now operates on. The multi-chain expansion broadens the addressable market but also multiplies the competitive surface.

The approximately 86% decline from ATHATHAll-Time High. The highest price a token has ever reached. ATH is usually quoted as a reference point for how far the current price has fallen (or risen) since the peak.Like the record lap time on a racetrack. It tells you what the car has been capable of at its absolute best, not what it will do today. Whether that record gets broken again depends on conditions that may or may not come back.ATH in the glossary → reflects market scepticism about the sustainability of the agent launchpad model. The buyback-and-burn and Unicorn launch system are structural improvements, but they don’t change the underlying dependency on narrative-driven speculation. If agent utility doesn’t improve beyond speculative token trading, these mechanisms defer the problem rather than solve it.

My position

I hold VIRTUAL. The multi-chain expansion, buyback-and-burn programme, and Unicorn launch system show a team actively iterating on the platform’s weaknesses. These are meaningful improvements from a year ago. The centralised inference engine remains the fundamental problem, and the revenue model is still narrative-dependent. Both are material risks I’m holding with open eyes. I would reassess if the GAME engine were open-sourced or made permissionless, and if revenue collapsed for a second consecutive cycle without recovery.

Freedom Score: 38/100

Virtuals Protocol scores 38/100 (F grade). Full methodology at Freedom Score Methodology.

Infrastructure decentralisation (5/20): The GAME inference engine runs on centralised hosted servers operated by Virtuals Protocol, constituting a significant single point of failure. AI agents depend entirely on Virtuals-hosted LLMs (Llama, DeepSeek, Qwen) with no permissionless node operation for inference. On-chain components (token contracts, governance, ACP settlement) inherit Base L2 decentralisation, which itself depends on a centralised sequencer operated by Coinbase. Agent data (personality, memory, voice, visuals) is stored on Virtuals-hosted infrastructure, not decentralised storage. The SDK offers local development but production deployment requires Virtuals-hosted endpoints.

Governance decentralisation (6/20): veVIRTUAL voting exists: any holder can vote, wallets with 0.10% of supply can propose, and votes run 72 hours with a 25% quorum. It doesn’t control upgrades, though. A single externally owned account can upgrade the core contracts with no timelock, so Virtuals’ claim that veVIRTUAL holders govern protocol upgrades is graded inflated on our integrity board. The team also retains significant influence: the 0.10% proposal threshold is high, the 35% treasury provides substantial team-aligned resources, and the founding team controls core infrastructure decisions.

Token distribution fairness (7/15): 60% of tokens went to public circulation via IDOs (Fjord Foundry LBP, Enjinstarter, PAID Network) in December 2021, a relatively fair public distribution. 5% to liquidity pools. The 35% treasury controlled by DAODAODecentralised Autonomous Organisation. A way to coordinate decisions and manage a treasury using token-weighted voting instead of a traditional company structure. Token holders propose and vote on changes directly.Like a shareholder-run company where every shareholder can vote on every decision, the votes are public, and the company can't do anything the shareholders don't approve. The coordination is messier than a normal company but nobody has unilateral control.DAO in the glossary → represents a large pool of team-aligned tokens, even with the 10% annual emission cap. The original PathDAO IDOIDOInitial DEX Offering. A token launch mechanism where the initial sale happens directly on a decentralised exchange, with built-in liquidity pool funding. IDOs replaced ICOs as the standard launch model for many projects after 2018.Like opening a new shop directly inside an existing market hall instead of running a separate sale event. The shop is open, the market provides foot traffic, and customers can walk in and trade from day one.IDO in the glossary → at $0.015 gave early investors a 338x ATH return. All tokens are fully unlocked. No VC allocation.

Censorship resistance (4/15): The platform can effectively censor agents. GAME Cloud is a hosted service with terms of service. The team controls which LLMs are available, can modify agent deployments, and controls the Agent Runner infrastructure. Agent tokens trade on Uniswap V2, but a single Virtuals wallet can blacklist any recipient of an AgentTokenV4 token, including its trading pair, which would halt every sell (see the AgentTokenV4 section). The AI inference layer is fully centralised too. An agent without GAME inference is just a token with no utility. Smart contracts on Base are upgradeable, allowing the team to modify contract behaviour.

Data sovereignty (5/15): Agent personality data, voice models, visual assets, and long-term memory are stored on Virtuals-hosted infrastructure. Users do not have self-custody of their agent data. Token holdings are self-custodied on-chain, but the intellectual property and functional components of agents are platform-hosted. The Immutable Contribution Vault stores approved contributions on-chain as ERC-1155 tokens, providing some data sovereignty for contribution records. Actual AI modelModelA trained neural network that takes inputs (text, images, audio) and produces outputs (more text, classifications, generated content). In DeAI the model is the thing that actually does the work.Like a very experienced apprentice who has spent years watching thousands of masters make furniture. They can't explain how they know when a joint is right, but they can make a chair that looks and functions like a Chippendale. The training is invisible. The output is what matters.Model in the glossary → weightsParametersThe internal numbers (weights and biases) inside a neural network that get adjusted during training. A 70-billion-parameter model has 70 billion adjustable internal numbers encoding everything it has learned.Like the synapses in a human brain. Each parameter is a tiny dial that gets nudged a little during training. With enough dials, the network can represent surprisingly complex patterns. The total parameter count is roughly how much "brain" the model has.Parameters in the glossary → and inference data remain on Virtuals servers.

Open source and transparency (11/15): Strong open-source presence with over 40 public GitHub repositories including protocol-contracts, virtuals-python SDK, openclaw-acp, and acp-node. Smart contracts verified on Basescan. Four security audits completed with reports published. GAME SDK is open source. Governance operates transparently via on-chain voting. However, the GAME Cloud engine (core inference infrastructure) is proprietary and hosted. HuggingFace presence is minimal (one model). Key operational components remain opaque.

Path to improvement

Three changes would materially increase Virtuals’ score:

1. Open-source or decentralise the GAME engine. The centralised inference layer is the single biggest freedom deficit. Either open-sourcing the GAME Cloud so anyone can run inference nodes, or implementing a permissionless provider network for agent inference, would address infrastructure, censorship, and data sovereignty scores simultaneously. This is the highest-impact change available.
2. Implement agent quality curation without censorship. The 18,000-plus agent count Virtuals markets is meaningless if the majority are inactive or spam. The Unicorn launch system and buyback-and-burn programme are steps in the right direction, but a broader reputation system, community curation layer, or minimum activity threshold for visibility (not existence) would improve platform quality without compromising permissionless creation.
3. Decentralise agent data storage. Moving agent personality data, memory, and assets from Virtuals-hosted infrastructure to decentralised storage (IPFS, Arweave, or similar) would give agent creators true ownership of their agents’ intellectual property. Currently, “owning” an agent means owning a token; the actual agent lives on Virtuals’ servers.

Returns Score: 66/100

VIRTUAL scores 66/100 (C grade). Full methodology at Returns Score Methodology.

Token utility (16/20): VIRTUAL has the strongest utility mechanics in the AI agent space. Every agent token on the core launchpad is paired against VIRTUAL in its DEX pools, creating mandatory demand that scales with trading activity across Base, Solana, and Ronin. Agent creation costs 100 VIRTUAL. The veVIRTUAL locking mechanism (up to two years) grants governance voting, daily Virgen points, and airdropAirdropDistributing tokens for free to eligible wallets, usually to reward early users, bootstrap a community, or decentralise token ownership away from a small group of insiders at launch.Like a supermarket handing out free samples to people who already shop there. The samples cost the supermarket nothing to print. The goal is to convert casual shoppers into loyal customers by giving them something tangible to talk about.Airdrop in the glossary → eligibility. The Unicorn conviction-based launch system replaced the original points-based Genesis model, introducing anti-sniper mechanics and valuation milestones. These are live, functioning mechanisms processing on-chain volume across multiple chains. The deduction reflects utility still being heavily tied to speculative agent launches rather than productive AI output.

Value accrual (12/20): The protocol captures on-chain fees, roughly $75M cumulative (DeFiLlama, October 2026) from 1% trading fees on agent tokens, agent creation fees and inference payments, and at peak over $1M a day flowed through it. About $42M of that stays with the protocol after creator payouts, and none of it reaches VIRTUAL holders: DeFiLlama tracks no holder revenue at all. The January 2025 buyback-and-burn spent approximately 13M VIRTUAL to buy and permanently burn 25 agent tokens, so it accrued to those agent tokens and left VIRTUAL’s supply unchanged. The core-launchpad base pair ties some VIRTUAL demand to agent trading, though not every dollar of it, since Occupy, Virtuals’ own launchpad, pairs its tokens with tokenised stocks. With no mechanical link from revenue to the token and agent tokens still trading like memecoins, the score sits in the speculative-accrual band.

Supply dynamics (12/20): The fixed 1B supply with all tokens fully unlocked is clean: no cliff events and no vesting schedules. The dilution that remains is the 35% treasury, which may release up to 10% a year for three years with governance approval. The no-VC, all-public-IDO distribution (60% public circulation, 35% treasury, 5% liquidity) is notably fair by industry standards. No team allocation exists. The January 2025 buyback-and-burn programme introduced deflationary pressure on agent tokens, though it burned no VIRTUAL. The largest burns targeted GAME (1.6M tokens), CONVO (1.1M tokens), and AIXBT (880K tokens). However, the 35% treasury remains a concentration concern, and the launchpad dynamics still fragment liquidity across thousands of agent token pools.

Revenue sustainability (16/25): The protocol generated on-chain-verifiable revenue, not token emissions disguised as yield, and the independent figure has grown: DeFiLlama tracks roughly $75M in cumulative fees (October 2026), up from the ~$39.5M we cited at the early-2025 peak, with about $16M in the trailing year across Base, Robinhood Chain, Solana and Ethereum, of which the protocol keeps about $5M. The multi-chain launchpad position with 18,000+ agents and $8B+ cumulative DEX volume (per Virtuals’ own reporting) demonstrates market fit for the tokenisation infrastructure. The ACP, Revenue Network and the Eastworld Labs robotics vertical broaden the base. The devastating counterpoint remains: the 97% revenue collapse within two months in early 2025 demonstrated that revenue depends on narrative-driven hype cycles, and the trailing year is a fraction of the peak run-rate. Cumulative revenue has grown, but the trailing year has fallen from about $20.7M to about $17M, so the extra point given in August for growth is withdrawn.

Liquidity and access (10/15): VIRTUAL is listed on Binance, Coinbase, OKX, Bybit, Kraken, and most major exchanges. The multi-chain deployment across Base, Solana, and Ronin materially reduces the bridging friction that was a weakness when the protocol was Base-only. The approximately 86% decline from ATH means strong liquidity coexists with substantial realised losses for most holders. The Unicorn anti-sniper tax system (99% decreasing to 1% over 98 minutes) adds launch-specific liquidity protection that the original bonding curve lacked.

Path to improvement

Three changes would materially increase Virtuals’ returns score:

1. Demonstrate cycle-independent revenue. The buyback-and-burn and community rewards programmes are positive steps, but the 97% revenue collapse from early 2025 remains the defining data point. Revenue that persists through a market downturn (from productive agent activity, ACP transaction fees, or enterprise use cases) would fundamentally change the investment thesis. The multi-chain infrastructure is in place; the revenue model needs to prove resilience.
2. Publish transparent agent quality metrics. The 18,000+ agent count is a vanity metric when the vast majority are inactive. Publishing daily active agents, inference calls, ACP transactions, and revenue per agent would let the market distinguish between quantity and quality. The Unicorn launch system is a start, but the market needs data to price conviction accurately.
3. Decentralise or open-source the GAME inference engine. The centralised inference layer is both a freedom deficit and a returns risk. If Virtuals’ servers go down, every agent stops functioning and the token’s utility evaporates. A permissionless inference provider network would reduce single-point-of-failure risk, which is ultimately a returns consideration as much as a decentralisation one.

Score change log

Date	Score	Change	Reason
2026-10-13	Editorial	N/A	The January 2025 buyback paragraph now says it burned agent tokens rather than VIRTUAL, matching the supply section. Scores unchanged.
2026-10-07	Score	Returns 66 held	Revenue Sustainability held at 16: kept revenue fell more than pump.fun’s through the same market, and the July rebound was one month.
2026-10-07	Data	N/A	DeFiLlama now books only the treasury share as revenue (creator payouts split out on 5 Oct), so the headline is cumulative fees, ~$75M, of which the protocol keeps ~$42M; trailing year ~$16M fees, ~$5M kept. Scores held pending review.
2026-10-06	Data	N/A	”~13M VIRTUAL burned” and “deflationary” removed from the description, verdict, pillars and pros: that VIRTUAL bought agent tokens to burn, and VIRTUAL supply is still exactly 1B on chain. Tokenomics now opens on demand, value capture and treasury dilution. Scores unchanged.
2026-10-04	Editorial	N/A	Product-in-practice sweep: removed the unsourced claim that most late-2025 agent-commerce volume was gamed, as no named analysis could be found, and dated the monthly rewards subsidy to April 2026. Scores unchanged.
2026-10-04	Data	N/A	Added Virtuals’ launch on Arc (17 Sep 2026, per its X post) and the Arc and Robinhood Chain VIRTUAL contracts its docs now list. Scores unchanged.
2026-09-26	Score	Freedom 39 → 38 (F held)	Censorship Resistance 5 → 4: one Virtuals wallet can blacklist any AgentTokenV4 recipient, including the trading pair, so agent-token trading is not censorship-resistant.
2026-09-27	Data	N/A	Correction: the token owner was called “a single key”; it’s an externally owned account, which a Fireblocks-style MPC wallet (Virtuals’ stated 2-of-3 setup) also looks like on-chain. Scores unchanged.
2026-09-27	Data	N/A	AgentTokenV4 section: the owner wallet’s tax-rate history from Base (57 calls since October 2024, cuts and raises, one V4 change). Scores unchanged.
2026-09-26	Data	N/A	Added an AgentTokenV4 section from our Base reads: one Virtuals wallet owns the tokens and can change the 1% tax without a cap or blacklist the pool; no V4 audit found.
2026-09-25	Score	Returns 68 to 66 (C held)	Value Accrual 14 to 12: Occupy pairs its tokens with tokenised stocks rather than VIRTUAL, and DeFiLlama tracks no holder revenue, so agent volume doesn’t reliably create VIRTUAL demand. Token Utility holds at 16.
2026-09-25	Data	N/A	The VIRTUAL base-pair claim now applies to the core launchpad only, since Occupy tokens pair with tokenised stocks; the January 2025 buyback burned agent tokens, so it no longer counts as accrual to VIRTUAL. Scores unchanged.
2026-09-24	Score	Freedom 42 to 39 (D to F); Returns 69 to 68	Governance 9 to 6: a single EOA upgrades the core contracts with no timelock. Revenue Sustainability 17 to 16: the trailing year fell, reversing August’s growth point.
2026-09-24	Data	N/A	Trailing-year revenue refreshed from ~$20.7M to ~$18M (DeFiLlama, September 2026); added the single-EOA, no-timelock upgrade path on the AgentNftV2 registry; repo count and the Returns 69/100 prose line corrected. Scores unchanged.
2026-09-24	Data	N/A	The 18,000+ agent count is Virtuals’ own multi-chain prototype tally, a different metric from our Base census of graduated agents; every mention now attributes it. Scores unchanged.
2026-09-24	Data	N/A	Added Occupy (September 2026): its tokens pair with tokenised stocks, and Virtuals’ published fee split sends 40% to a stock treasury, 30% to creators and 30% to Virtuals Protocol with no stated use for VIRTUAL. Scores unchanged.
2026-09-23	Data	N/A	Integrity records re-cut: 6 claims now quote Virtuals Protocol’s own pages verbatim and 14 are our own measurements. Scores unchanged.
2026-09-05	Data	N/A	§8e sweep correction. Replaced an unsourced “18,000 launched, fewer than a hundred above $1M” with our own registry census, bound live: 1,232 graduated agents, 85% dead or near-dead, 6 above $1M. The 18,000 figure is multi-chain cumulative prototypes, which our own data file flags as a different metric.
2026-08-17	Data	N/A	Exchange ticker count corrected from 100 to 457. CoinGecko serves 100 tickers per page and publishes no total, so the original unpaginated read returned the page size and it read like a measurement. The reader now paginates and flags the bound.
2026-08-15	Data	N/A	Survivorship refreshed (2026-06-09 index to 2026-08-10): dead or near-dead graduated agents 71% to 84%, and agents turning over $1M or more in 30 days 25 to 3, while the graduated count barely moved. The “productive society” thesis verdict stays inflated, on stronger evidence.
2026-08-13	Data	N/A	Sourced the “$470M Agentic GDP” claim to Virtuals’ own press release of 12 February 2026, so it is no longer unattributed. Verdict stays unverifiable: aGDP is Virtuals’ own coinage with no published methodology and no independent series to check it against.
2026-08-13	Data	N/A	Re-pointed all 17 Virtuals claim sources from the bare whitepaper.virtuals.io homepage to the page carrying each claim, so a reader can trace every verdict. The tokenomics page confirms the fixed 1B supply “minted without future inflation” and the 60/5/35 public/liquidity/treasury split verbatim.
2026-08-13	Data	N/A	The “$470M Agentic GDP” claim is now marked unsourced: Virtuals made it in marketing but we never recorded where, and pointing it at a plausible whitepaper page would be worse than saying so. Verdict stays unverifiable.
2026-08-03	Returns	68 → 69	Revenue Sustainability 16→17. DeFiLlama cumulative revenue ~$73M (Base + Robinhood Chain + Solana + Ethereum), up from the stale ~$39.5M early-2025 figure; base broadened (Revenue Network, Eastworld Labs). C grade held.
2026-08-03	Data	N/A	Refreshed the cumulative-revenue figure across body, snapshot, pillars and dimensions from ~$39.5M (early 2025) to ~$73M (DeFiLlama, Aug 2026; Robinhood Chain now a tracked revenue chain). Added the July 2026 LayerZero→Chainlink CCIP cross-chain migration (post $292M LZ exploit).
2026-07-06	Editorial	N/A	Added a “The product in practice” section (per specs/product-comparison-framework.md): substantive launchpad/GAME/ACP infrastructure but speculation-led (most agent-commerce volume found gamed; fewer than 100 of 18k+ agents above $1M). No score change.
2026-05-27	Editorial	N/A	EconomyOS launch (16 May 2026) recorded: managed-agent suite giving onchain agents email inboxes, non-custodial wallets, virtual Visa cards, and identity layer for autonomous Web2/Web3 commerce. Self-reported metrics (17K agents, ~$479M aGDP, 1.77M jobs per Crypto Briefing 16 May) noted but not adopted into prose pending DeFiLlama / Token Terminal cross-check. DeFiLlama 30d fees remain $427K against the self-reported numbers, gap unchanged. No score change.
2026-05-02	Editorial	N/A	Added ERC-8004 (Trustless Agents) registration rollout to all graduated ACP agents (April 2026). Aggregator-cited cumulative revenue figures (~$60M) noted but not adopted pending DeFiLlama/Token Terminal cross-check. No score change.
2026-04-06	Data	N/A	Chain expansion (Arbitrum, XRP Ledger). Revenue Network, Eastworld Labs, three-tier launch system detail added. Agent GDP updated.
2026-03-12	Returns	64 → 68	Multi-chain expansion (Solana, Ronin), buyback-and-burn programme (~13M VIRTUAL), Unicorn launch system, community rewards. Revenue Sustainability and Supply Dynamics improved.
2025-03-03	Both	N/A	Initial publish. Freedom 42/100, Returns 64/100.