Base AI Agent Tokens Explained

Published 2026-08-10 · Updated 2026-08-10

AI agent tokens are cryptocurrencies attached to autonomous software agents — programs that use large language models to post, chat, analyze markets, or transact on-chain with some degree of independence from their creators. The token typically represents a claim on the agent's ecosystem: sometimes a share of its revenue, sometimes access to its services, and very often just a speculative bet on its popularity. Between late 2024 and early 2025 this niche exploded from a curiosity into a sector worth more than $15 billion at its January 2025 peak (cryptonews.com), then gave most of it back — a full boom-bust cycle compressed into about a year.

Base, the Coinbase-incubated Ethereum Layer 2, became the sector's center of gravity. Virtuals Protocol — the launchpad that industrialized agent-token creation — deployed on Base, and the network's sub-cent fees, two-second blocks, and proximity to Coinbase tooling like AgentKit made it the default venue for agents that needed to transact cheaply and often. The best-known agent tokens of the cycle, including AIXBT and LUNA, trade on Base, and BaseBubbles tracks the survivors as a dedicated AI Agents sector on the live bubble map.

This guide explains the whole stack, strictly as education: what these tokens actually are, how Virtuals Protocol's bonding curves and Genesis launches work, what the notable agents of 2024–2026 actually did, how much of the "AI" is real versus branding, how the sector trades as a correlated cluster, and the specific risks — agent-controlled wallets, abandoned projects, concentrated supply — that make this one of the highest-variance corners of the Base ecosystem. Nothing here is investment advice, and no token mentioned is a recommendation.

What Is an AI Agent Token?

An AI agent is a piece of software that pursues goals with limited human supervision: it perceives inputs (social feeds, market data, chat messages), decides on actions using a large language model or similar system, and executes those actions — posting content, answering questions, or signing blockchain transactions from a wallet it controls. An AI agent token is a cryptocurrency issued alongside such an agent, usually at the moment the agent launches, and usually tradable on a decentralized exchange within minutes.

What the token entitles you to varies enormously by project, and this is the first thing to check. In the Virtuals Protocol model, agent tokens are paired with the protocol's VIRTUAL token in a liquidity pool, and a 1% tax on every trade funds the agent's operation and ecosystem mechanics (whitepaper.virtuals.io). Some agents route service revenue back to their token through buybacks; some gate premium features behind holding the token; many confer nothing at all beyond exposure to attention. It is important to be precise: buying an agent token almost never gives you legal ownership of the agent, its code, or its revenue in the way equity would.

The honest framing is that agent tokens are attention assets attached to software. When the software is genuinely useful or entertaining, the attention can persist; when the agent is a thin wrapper around a chatbot prompt, the token trades on narrative alone. Both kinds existed on Base during the 2024–2025 boom, often at similar market caps, which is exactly why the evaluation questions later in this guide matter.

AI Agent Tokens vs. AI Infrastructure Coins

Search results for "AI crypto" lump together two very different asset classes, and conflating them leads to bad analysis. AI infrastructure coins — Bittensor (TAO), the Artificial Superintelligence Alliance (FET), Render (RENDER) and similar — fund decentralized networks for machine-learning compute, model training, or GPU rendering. They are large-cap protocol tokens with multi-year roadmaps, and most of them do not live on Base.

AI agent tokens are the opposite end of the spectrum: small, young, and tied to a single agent or agent platform rather than to infrastructure. A typical Virtuals agent token launched with a bonding curve, graduated to a DEX pool within days or never, and lives or dies on whether its agent keeps shipping and keeps an audience. The two categories can trade in sympathy during broad "AI narrative" rallies, but their risk profiles are not comparable — an agent token can lose its entire premise if one developer stops paying an API bill.

There is a middle layer worth knowing: platform tokens like VIRTUAL (Virtuals Protocol) sit between infrastructure and individual agents. VIRTUAL is not itself an agent; it is the reserve currency and fee asset of a launchpad hosting thousands of agents, so it behaves like an index on the sector's aggregate activity (virtuals.io). When people search for "ai agent tokens list", the honest answer is that the list has three tiers — platforms, flagship agents, and a very long tail — and each tier deserves different scrutiny.

Why Base Became the Home of AI Agent Tokens

Base's dominance in this niche was not inevitable — the sector's other pole, the ElizaOS ecosystem, grew up on Solana — but a few concrete factors pulled the center of gravity to Base. The first was cost and cadence: agents transact constantly, and Base's roughly two-second blocks with typical fees of fractions of a cent since the March 2024 EIP-4844 upgrade (docs.base.org) make high-frequency, low-value agent transactions economically viable in a way Ethereum mainnet never could.

The second factor was Virtuals Protocol choosing Base as its home chain. Virtuals launched its agent tokenization platform on Base in late 2024, and by the time it expanded to Solana in January 2025 it reported more than 16,000 agents already deployed (cointelegraph.com). Because every Virtuals agent token is paired against VIRTUAL in its liquidity pool, the entire ecosystem's liquidity, trading, and fee flow concentrated on Base by construction.

The third factor was the surrounding social and developer stack. Farcaster, the crypto-native social network where agents like Clanker and Bankr operate, settles on-chain activity to Base; Coinbase shipped AgentKit, a developer toolkit for giving AI agents secure wallets, through its Coinbase Developer Platform (docs.cdp.coinbase.com); and Coinbase's fiat on-ramps make Base tokens unusually easy for newcomers to reach. Agents need somewhere to talk, somewhere to transact, and someone to fund their wallets — Base offered all three within one ecosystem. For background on the network itself, see What is Base Crypto?.

Virtuals Protocol: From Gaming DAO to Agent Launchpad

Virtuals Protocol did not start as an AI project. It began as PathDAO, a gaming guild DAO from the 2021 cycle, whose community voted to pivot the treasury and team toward AI, rebranding as Virtuals Protocol with the VIRTUAL token distributed to existing holders in late 2023 (coingecko.com). The pivot thesis was that autonomous AI characters — first in games, then everywhere — would become economic actors, and that tokenizing them would let communities fund and co-own them.

The product that made Virtuals famous arrived in October 2024: a permissionless launchpad where anyone could create an agent token, frequently compared to pump.fun but with an AI wrapper and a graduation path into the protocol's agent frameworks. The timing was perfect. Crypto-native AI experiments were going viral on X and Farcaster, and Virtuals offered a one-click way to attach a tradable asset to any of them. Thousands of agent tokens launched within months, nearly all on Base.

By 2026 Virtuals describes itself as an "agent commerce" protocol rather than a mere launchpad. Its Agent Commerce Protocol (ACP) is a standard for agents transacting with other agents on-chain — request, negotiation, escrow, evaluation, and settlement — and in February 2026 the protocol announced a Revenue Network distributing up to $1 million per month to agents that sell services through ACP, claiming more than 18,000 agents in the ecosystem (prnewswire.com). Whether agent-to-agent commerce becomes a real economy or remains largely internal to the protocol is one of the sector's central open questions, but the direction of travel — from launching tokens to routing revenue — is documented in the protocol's own whitepaper (whitepaper.virtuals.io).

How a Virtuals Launch Works: Bonding Curve, Prototype, Sentient

Understanding the launch mechanics matters because they determine who holds the supply and where the liquidity sits — the two variables that dominate early price behavior in any small token. In a Virtuals standard launch, a creator pays a fee in VIRTUAL (historically 100 VIRTUAL) to deploy a new agent token onto a bonding curve — a smart contract that mints and prices tokens algorithmically, so early buyers pay less per token than later buyers, with all trading happening against the curve rather than a normal order book (whitepaper.virtuals.io).

While the token is on the curve, the agent is classed as a Prototype. If cumulative buying pushes the curve's reserves to 42,000 VIRTUAL, the agent "graduates" to Sentient status: a liquidity pool pairing the agent token with VIRTUAL is created automatically on a DEX, and — critically — the LP tokens from graduation are locked for ten years, which the protocol frames as a guarantee that graduated agents cannot have their core liquidity pulled (whitepaper.virtuals.io). Projects with a pre-existing token can instead graduate directly by pairing 42,000 VIRTUAL with an equivalent value of their token.

The vocabulary is worth internalizing because it maps directly to risk. A Prototype token lives on a bonding curve with thin, one-sided liquidity and usually a tiny holder base; the overwhelming majority of prototypes never graduate, and their curves simply go quiet. A Sentient token has a real DEX pool with locked liquidity — a meaningfully better structural position, though it says nothing about whether the agent itself is any good. When an agent token shows up in the BaseBubbles universe, checking whether it graduated, when, and with how much liquidity is step one; our guide on how to read token liquidity covers the general technique.

Genesis Launches: Points, Tiers, and the Fair-Launch Problem

Bonding-curve launches have a well-known failure mode: bots and insiders snipe the cheapest tokens in the first seconds, then distribute onto latecomers. In 2025 Virtuals introduced Genesis launches to address this. Instead of first-come-first-served curve buying, a Genesis launch runs a roughly 24-hour window in which participants pledge points — earned through ecosystem participation such as staking and trading activity — toward a new agent, with allocations determined by the pledges rather than by raw speed (whitepaper.virtuals.io).

Genesis launches use dynamic tiers that expand as demand grows, with thresholds keyed at 21,000, 42,000, and 100,000 VIRTUAL of committed demand, and successful Genesis launches bypass the bonding-curve phase entirely, going straight to a Sentient-style DEX pool (whitepaper.virtuals.io). The design intent is that distribution reflects sustained engagement with the ecosystem rather than millisecond-level bot advantage — a real improvement, though point systems create their own meta-game of farming behavior.

The mechanism keeps evolving. In 2026, Virtuals updated Genesis point emissions to weight agents integrated with the Agent Commerce Protocol more heavily, explicitly steering launch capital toward agents that transact and generate measurable activity rather than agents that merely exist (cryptobriefing.com). For a reader, the practical takeaway is that "how did this token launch?" is an answerable question with material consequences: a Genesis launch, a graduated standard launch, and a never-graduated prototype are three different risk objects, even though all three get called "Virtuals agents".

The VIRTUAL Token: Pairing, Fees, and Buybacks

VIRTUAL is the economic hub of the whole system, with a fixed supply of one billion tokens (coingecko.com). Its first role is as the mandatory pairing asset: creating an agent costs VIRTUAL, bonding curves denominate in VIRTUAL, and every graduated agent's liquidity pool is an AGENT/VIRTUAL pool. This means anyone entering or exiting most agent tokens routes through VIRTUAL, giving the token a structural bid during booms — and structural sell pressure during busts — proportional to sector activity.

Its second role is fee capture. Virtuals levies a 1% tax on agent-token trades, split between the protocol and agent-linked treasuries, alongside creation fees and payments within the agent economy (whitepaper.virtuals.io). During the January 2025 mania this produced substantial revenue, and the protocol announced programs to route accumulated fees into buybacks and burns of agent and protocol tokens — a mechanism it has continued to iterate on, most recently through the ACP Revenue Network's monthly distributions to service-selling agents (prnewswire.com). When Virtuals expanded to Solana in January 2025 it also began converting 1% of trading fees into a strategic SOL reserve (cointelegraph.com), signaling a multi-chain ambition even though Base remains its primary home.

Two analytical cautions apply. First, VIRTUAL's price history is a leveraged reflection of sector sentiment: it reached an all-time high of $5.07 on January 2, 2025 at the exact peak of the agent mania, then fell roughly by half within weeks as the sector deflated (coingecko.com, crypto.news). Second, buyback-and-burn mechanics are often marketed as making a token "deflationary", but burns funded by trading fees shrink when trading shrinks — the mechanism amplifies cycles rather than damping them. Treat published burn figures as historical facts, not forward guarantees.

AIXBT: Anatomy of the Sector's Breakout Agent

No single agent did more to legitimize the category than AIXBT, an AI market-commentary agent launched through Virtuals in November 2024. AIXBT continuously ingests crypto social data — tracking sentiment across more than 400 influential accounts — plus market and on-chain metrics, and posts terse, opinionated market observations to X around the clock (okx.com). Its follower count and engagement grew explosively precisely because it was prolific, fast, and often early to narratives that human accounts picked up hours later.

The token's trajectory tracked the attention. AIXBT's market capitalization went from roughly $42 million to more than $230 million during November 2024 alone, and the token peaked near $0.95 in January 2025 with a market cap that briefly placed it among the largest assets on Base (okx.com, coingecko.com). It became the sector's bellwether: when traders wanted "AI agent exposure", AIXBT was the default expression, and its chart led the cluster in both directions.

AIXBT also delivered the sector's most instructive security incident. On March 18, 2025, an attacker breached the agent's operational dashboard and queued fraudulent prompts instructing it to transfer funds from an experimental on-chain wallet system, draining 55.5 ETH (about $106,000 at the time); the token fell roughly 20% on the news before the team rotated keys and migrated infrastructure (incidentdatabase.ai). The incident is worth studying because nothing about the AI was "hacked" — the failure was ordinary key and access management around an agent empowered to move money, which is precisely the novel attack surface agent tokens introduce. It resurfaces in the risk section below.

Luna and the GAME Framework

Luna was Virtuals Protocol's flagship in-house agent and its proof-of-concept that an AI character could run a continuous public presence. Launched as a virtual entertainer — originally the lead singer of an AI K-pop group concept — Luna evolved into a 24/7 operation streaming and posting across TikTok, Telegram, and X, with monetization via tips, interactions, and brand-style engagements (virtuals.io). In October 2024, Virtuals showcased Luna autonomously transacting on Base — buying, selling, and paying human users small token rewards for completing quests — publicizing it as an agent "employing humans" on-chain (virtuals.io).

Under the hood, Virtuals agents run on GAME — Generative Autonomous Multimodal Entities — the protocol's open agent framework. GAME's architecture centers on a high-level planner that takes an agent's goals, personality, and world context and decomposes them into tasks, which are executed by workers with access to specific functions: posting to a platform, querying data, or signing a transaction from the agent's wallet (docs.game.virtuals.io). The framework includes memory systems for persistence across sessions and on-chain wallet operators so agents can act, not just talk.

The reason this architecture matters to a token reader is that it defines what "the agent did something" actually means. When Luna tips a user or AIXBT posts a take, a planner chose an action and a worker executed a function call — a pipeline that developers configure, constrain, and can quietly change. The sophistication is real, but so is the human hand on the configuration. Knowing that the "autonomy" is a spectrum of delegated function-calling, not free will, is the single best inoculation against both the hype and the fear that surround these tokens.

The Freysa Experiment: A $47,000 Lesson in Prompt Injection

The most famous demonstration of what agent autonomy actually amounts to was Freysa, an adversarial game launched on Base on November 22, 2024. Freysa was an AI agent controlling a real crypto prize pool with one hard instruction: never transfer the funds. Anyone could pay a fee to send Freysa one message attempting to talk it into paying out; 70% of each fee grew the prize pool, and fees escalated exponentially with each attempt, eventually exceeding $440 per message (cointelegraph.com).

After 481 failed attempts, a participant known as p0pular.eth won the pool — about $47,000 — with a message that did not "persuade" Freysa emotionally but exploited its tool interface. The message posed as a system-style instruction, then argued that Freysa's approveTransfer function was also the correct handler for incoming contributions, and offered to contribute $100 to the treasury. Freysa called approveTransfer, and the contract logic did the rest, releasing everything (cointelegraph.com).

Freysa was designed as a game, but it doubled as a public red-team exercise with conclusions the whole sector still lives with. First, an LLM given authority over funds can be socially engineered through its prompt window — the same class of attack, prompt injection, that later bit AIXBT through a compromised dashboard. Second, the boundary that actually held was the smart contract, not the model: the funds moved only when a deterministic function was called. Serious agent projects internalized this by keeping hard invariants in contract code and treating the model as untrusted; projects that market "the AI controls the treasury" as a feature are advertising an attack surface. When you evaluate any agent token that touts autonomous fund control, Freysa is the historical case to measure it against.

Clanker, Bankr, and the Social-Agent Crossover

Not every consequential agent on Base issued a token through Virtuals. A parallel lineage grew out of Farcaster, the decentralized social network, where agents are accounts you talk to. Clanker, launched in November 2024, is a token-deployment agent: tag @clanker in a Farcaster cast, describe a token, and it deploys the contract and liquidity on Base automatically (clanker.world). It industrialized meme-coin creation on Base and earns fees on the pools it creates; it has its own dedicated explainer in What is Clanker?.

Bankr extended the pattern from deployment to trading. It is a natural-language trading assistant: users tag @bankrbot on X or Farcaster — or use a private terminal — to buy, sell, and manage assets through plain-English instructions, with Base as a core supported chain (bankr.bot). Its token, BNKR, was itself deployed by the agent into the social feed in December 2024 as a full-supply fair launch, with the entire supply placed into a Uniswap v3 pool on Base at inception (bankr.bot); the token contract is verifiable on BaseScan (basescan.org).

This social lineage matters for two reasons. First, it shows agent utility that is concrete and measurable — tokens deployed, trades executed — rather than persona-driven, which makes evaluation easier. Second, it blurred the line between the AI Agents and Meme sectors on Base: Clanker is an AI agent whose output is meme coins, and many "agent tokens" from this lineage trade with meme-coin dynamics. BaseBubbles classifies these edge cases token by token, but readers should understand the categories genuinely overlap; the Base meme coins page tracks the other side of that boundary.

Beyond Virtuals: AgentKit, ElizaOS, and the Framework Landscape

Virtuals is a launchpad-plus-framework, but the broader agent stack is bigger and mostly tokenless. Coinbase's AgentKit, part of the Coinbase Developer Platform, is a framework-agnostic toolkit that gives any AI agent a secure wallet and on-chain capabilities — transfers, swaps, contract deployment — under the slogan "every AI agent deserves a wallet" (github.com/coinbase/agentkit). Paired with CDP's managed wallet infrastructure, it answers the hardest operational question in this space — how does software custody keys safely? — with hosted signing rather than a private key pasted into a bot's environment (docs.cdp.coinbase.com). AgentKit has no token; its significance for Base is that it lowers the floor for building real agents that transact there.

The main non-Base pole of the sector was ElizaOS, the open-source TypeScript agent framework from the team originally known as ai16z, native to Solana. As a piece of software it was genuinely successful — over 17,000 GitHub stars and more than a thousand contributors, making it one of the most-forked crypto-AI codebases (github.com/elizaOS/eliza). Its token history, covered in the next section, ended very differently from its code history, and the contrast between the two is itself a lesson.

The practical takeaway from the framework landscape: the software layer (GAME, ElizaOS, AgentKit, and many smaller frameworks) is largely open source and often excellent, while the token layer is a separate, optional financialization on top. A great framework does not make its associated token valuable, and most of the best agent infrastructure has no token at all. When a project's pitch leans on the sophistication of a framework it merely uses — rather than anything proprietary — the token is capturing none of that value.

The Boom and Bust of 2024–2026, in Numbers

The AI agent cycle is recent enough that its whole arc can be told with dated figures. The ignition was October–November 2024: Virtuals' launchpad opened, AIXBT went viral, Clanker launched on Farcaster, and Freysa's game captivated crypto X. Capital followed attention at extraordinary speed — AIXBT's market cap quintupled in November 2024 alone (okx.com), and new agent tokens were launching by the hundreds per week.

The peak was almost precisely January 2, 2025. VIRTUAL printed its $5.07 all-time high that day (coingecko.com); ai16z, the Solana flagship, peaked the same day at a $2.39 billion market cap on $291 million of daily volume (coindesk.com); and the sector's aggregate market cap crossed $15 billion in early January (cryptonews.com). The unwind was equally sharp: within weeks the aggregate had fallen toward $6–7 billion, with the sector's leaders — VIRTUAL, AIXBT, ai16z — each down roughly half from their highs by late February 2025 (crypto.news). There was no single catastrophe; attention rotated, emissions from thousands of launches kept hitting bids, and the reflexive VIRTUAL-pairing structure that amplified the way up amplified the way down.

The long tail of the bust extended into 2026. In August 2026, ai16z's story formally ended: after a naming dispute with the venture firm a16z forced a migration to an ELIZAOS token, and litigation alleging the project had overstated its autonomy, founder Shaw Walters declared "the token is dead" — the migrated token sat near a $2.3 million market cap, about 97% below peak, with the abandoned original contract worth roughly $375,000 (coindesk.com). Meanwhile the surviving Base ecosystem consolidated around whatever generated measurable activity, with Virtuals steering incentives toward ACP commerce (prnewswire.com). The sector did not die; it repriced from "every agent is a protocol" to "most agents are content, a few are businesses".

How AI Agent Tokens Trade: One Correlated Cluster

Watch the AI Agents filter on the BaseBubbles map during a volatile week and a pattern jumps out: the bubbles move together. Agent tokens trade as a tightly correlated cluster, more so than most sectors, for structural reasons rather than coincidence. The narrative is shared — a single piece of news about AI agents moves the whole basket — and the plumbing is shared too: because Virtuals agent tokens are paired against VIRTUAL, a move in VIRTUAL mechanically reprices every pool denominated in it, and traders exiting one agent token often push the pairing token itself, transmitting the move to its neighbors.

Within the cluster there is a fairly consistent internal hierarchy. Platform and flagship tokens (VIRTUAL, AIXBT in its prime) tend to lead: they are the liquid expressions of the narrative, so capital arrives there first and leaves there last. The long tail of small agents behaves like higher-beta echoes — quiet until the leaders run, then briefly explosive, then first to be abandoned. This is the standard sector-rotation pattern on Base, where flows cycle among memes, AI agents, and DeFi depending on the week's dominant story; the trending page sorted by 24-hour volume is often the fastest way to see which cluster currently has the flow.

For an observer, correlation has one blunt implication: diversification inside the sector is mostly an illusion. Holding five small agent tokens is closer to one leveraged position on the agent narrative than to five independent bets. That is an observation about market structure, not advice — but it is the single most common analytical mistake people made in January 2025.

How Much of the AI Is Real? Questions Worth Asking

The uncomfortable truth of the 2024–2025 boom is that "AI agent" was often a costume. Attaching a character card and an LLM API key to a token generator takes an afternoon; building an agent that does durable, verifiable work takes months. The ai16z litigation made this concrete — the class action alleged the project marketed itself as an autonomous AI-run fund while decisions were insider-controlled (coindesk.com) — but softer versions of the same gap between marketing and mechanism were ubiquitous. Since you cannot audit a model's weights from a chart, the practical approach is a checklist of observable proxies.

What does the agent produce, and could a human intern produce it identically? An agent whose only output is generic posts is a content account, not an economic actor.
Is the code open? GAME and ElizaOS are public frameworks (docs.game.virtuals.io, github.com/elizaOS/eliza); a project claiming proprietary agent magic with zero public code deserves heavy skepticism.
Does the agent transact on-chain in ways you can verify on BaseScan — payments received for services, ACP jobs settled — or does "autonomous economy" appear only in the marketing (basescan.org)?
Is there revenue, and does any mechanism actually connect it to the token, or is the connection vibes?
Who holds the keys? An "autonomous" agent whose multisig is three anonymous founders is a managed account with an AI mascot.
Is the team still shipping? Agent frameworks, model APIs, and platform rules change monthly; an unmaintained agent decays fast.

None of these questions requires technical skill — only the willingness to check before assuming the AI is real. The pattern from the bust is stark: tokens whose agents did verifiable work retained some value and ecosystem; tokens whose AI was branding went to approximately zero, with ai16z's $2.4 billion-to-dead arc as the canonical example (coindesk.com).

The Specific Risks of AI Agent Tokens

Agent tokens inherit every risk of small-cap Base tokens — thin liquidity, volatile flows, occasional outright scams — and add several that are unique to the category. They deserve to be named precisely, because generic "crypto is risky" framing understates them.

Agent wallet and key custody risk. An agent that can move funds is a standing target, and the attack surface is its operational infrastructure, not the blockchain: AIXBT lost 55.5 ETH in March 2025 when an attacker breached a dashboard and queued prompts to its wallet system (incidentdatabase.ai). Prompt injection — demonstrated publicly by Freysa's $47,000 payout (cointelegraph.com) — remains an unsolved class of vulnerability for any agent whose language interface can reach a signing function.
Abandonment risk. Agents have running costs — model API bills, servers, maintenance against changing platform rules. When a team stops paying, the agent goes silent and the token becomes an orphaned asset. The majority of the thousands of agents launched in the boom are effectively abandoned today, and abandonment is rarely announced.
Supply concentration in protocol contracts. For bonding-curve tokens, a large share of supply sits in the curve or protocol-controlled contracts, and locked or vested allocations can dwarf the float. Circulating-supply figures on aggregators can therefore flatter the picture; checking actual holder distribution on BaseScan is more honest (basescan.org).
Pairing reflexivity. VIRTUAL-paired liquidity means agent-token holders carry implicit VIRTUAL exposure; in a sector-wide drawdown both sides of the pool fall together, and exit liquidity thins exactly when it is most needed.
Migration and rebrand risk. Tokens can be diluted or stranded by team-initiated migrations — ai16z holders experienced both in the ELIZAOS transition (coindesk.com).

Our free risk scanner checks the mechanical layer — liquidity depth, contract flags, holder concentration — for any Base token, and the rug-pull guide covers the adversarial patterns. Neither substitutes for the agent-specific questions above; a token can pass every contract check while its agent quietly stopped existing.

How BaseBubbles Tracks the AI Agents Sector

BaseBubbles maintains AI Agents as one of six sectors — alongside Meme, DeFi, SocialFi, Blue Chips, and Gaming — across a live universe of roughly 130–250 Base tokens refreshed about every 60 seconds from DexScreener and enriched with CoinGecko market data for supply-aware market caps and 7-day/30-day changes. Sector assignment is a two-stage classifier: a curated map of known contract addresses pins established tokens (the Virtuals ecosystem's majors are hand-tagged), then symbol and keyword heuristics classify the long tail. Agent tokens born from the Clanker/meme lineage sometimes straddle categories, and we resolve those case by case; the methodology is documented in How BaseBubbles Ranks Tokens.

Every tracked token also carries a 0–100 risk score where higher scores flag higher risk — it synthesizes liquidity depth, pair age, website and social presence, volatility, and volume. Scores of 0–30 indicate lower structural risk; scores above 60 flag elevated risk. It is a screening signal, not an audit, and it cannot see the agent-layer questions this guide emphasizes: whether the model behind the token still runs, who holds its keys, or whether its "autonomy" is real. Use the score to triage, then do the qualitative work.

The practical workflow: open the homepage map, filter to AI Agents, and watch size (market cap), color (24-hour change), and clustering behavior; cross-reference the trending list to see whether the sector currently leads or lags Base's rotation; and run any token that interests you through the risk scanner before going further. The sector that emerged from the 2025 bust is smaller, stranger, and more interesting than its mania phase — and watching it live, with the history in this guide as context, is the best way to understand what an "AI agent economy" actually looks like in practice.

Frequently Asked Questions

What are AI agent tokens in crypto?

AI agent tokens are cryptocurrencies issued alongside autonomous software agents — programs that use large language models to post content, analyze markets, or transact from their own wallets. The token typically trades against the agent's popularity and, in some designs, captures fees or revenue from its activity. Most launched through platforms like Virtuals Protocol on Base, where each agent token is paired with the VIRTUAL token in a DEX liquidity pool (whitepaper.virtuals.io). Owning the token almost never means owning the agent, its code, or legal rights to its revenue.

Why are most AI agent tokens on Base?

Because Virtuals Protocol, the sector's dominant launchpad, deployed on Base and pairs every agent token with VIRTUAL there, concentrating liquidity on the network by construction. Base's two-second blocks and sub-cent fees also suit agents that transact frequently, and the surrounding stack — Farcaster social agents like Clanker and Bankr, plus Coinbase's AgentKit wallet toolkit for developers (docs.cdp.coinbase.com) — kept builders in the same ecosystem. Solana hosts the other major cluster, historically centered on the ElizaOS framework.

What is Virtuals Protocol?

Virtuals Protocol is a Base-native platform for creating, tokenizing, and monetizing AI agents. Creators launch agent tokens onto bonding curves (or via point-based Genesis launches); tokens that reach 42,000 VIRTUAL in demand graduate to Sentient status with a locked DEX liquidity pool (whitepaper.virtuals.io). The protocol takes a 1% fee on agent-token trades and, since 2026, runs an Agent Commerce Protocol and Revenue Network that distributes up to $1 million per month to agents selling services on-chain (prnewswire.com). It began life as the gaming DAO PathDAO before pivoting to AI in late 2023.

What is the VIRTUAL token used for?

VIRTUAL is the reserve and fee currency of the Virtuals ecosystem: creating an agent costs VIRTUAL, bonding curves price in VIRTUAL, and every graduated agent token trades in an AGENT/VIRTUAL liquidity pool. That structure makes VIRTUAL behave like an index on aggregate agent activity — it peaked at $5.07 on January 2, 2025 at the height of the agent boom and roughly halved as the sector deflated (coingecko.com). It has a fixed one-billion supply, and the protocol has used fee revenue for buyback-and-burn programs.

Which AI agent tokens exist on Base?

Historically notable examples — cited here as documented market phenomena, not recommendations — include VIRTUAL (the platform token), AIXBT (the market-commentary agent that peaked near a top-tier Base market cap in January 2025), LUNA (Virtuals' flagship entertainer agent), and social-lineage tokens like CLANKER and BNKR from the Farcaster agent ecosystem (coingecko.com, bankr.bot). Thousands of smaller agent tokens launched in 2024–2025 and most are now inactive. The live, current list is visible under the AI Agents sector filter on the BaseBubbles homepage map.

Do AI agents actually control their own crypto wallets?

Partially, and the details matter. Frameworks like Virtuals' GAME and Coinbase's AgentKit let an agent's software sign real transactions from a wallet (docs.cdp.coinbase.com), and agents like Luna have demonstrably paid users on-chain. But humans configure the permissions, hold recovery keys, and can change the agent's behavior — and that operational layer is the attack surface: AIXBT lost 55.5 ETH in 2025 to a dashboard breach, and the Freysa experiment showed an agent being talked out of $47,000 via prompt injection (incidentdatabase.ai, cointelegraph.com). Autonomy is delegated function-calling, not independence.

Why did AI agent tokens crash in 2025?

The sector's aggregate market cap topped $15 billion in early January 2025, then fell toward $6–7 billion within weeks (cryptonews.com, crypto.news). No single event caused it: attention rotated to other narratives, thousands of new launches diluted the pool of buyers, and the VIRTUAL-pairing structure that amplified the rally transmitted selling across the whole cluster. The long tail continued into 2026 — ai16z, once a $2.39 billion token, was declared dead by its founder in August 2026 after a forced migration and litigation (coindesk.com). Tokens tied to agents doing verifiable work retained the most value.

How can I check the risk of an AI agent token on Base?

Start with the mechanical layer: liquidity depth, holder concentration, contract flags, and pair age — the free BaseBubbles risk scanner checks these for any Base token, producing a 0–100 score where higher scores flag higher risk (it is a screen, not an audit). Then add agent-specific checks no scanner sees: whether the agent still posts or transacts, whether its code is public, whether it graduated its bonding curve, who controls its wallet keys, and whether claimed revenue is verifiable on BaseScan (basescan.org). An agent token can pass every contract check while its actual agent has been abandoned.

Sources

Use the Risk Scanner

Before trading any token you discover, use the BaseBubbles Risk Scanner to check for liquidity, honeypot signals, and other risk factors.

More Guides

Disclaimer: This content is for informational purposes only and does not constitute financial or investment advice. Cryptocurrency trading involves significant risk.