The AI Agent Landscape: What Actually Helps, and What's Just a Token

The word “agent” has been stretched to cover almost anything. A chatbot with a system prompt. A trading bot. A token on a launchpad with a cartoon avatar. A robot runtime. A privacy client for decentralised 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.Read more →. They wear the same label and earn opposite verdicts.

This is a map, not a buyer’s guide for tokens. The question is simple: which agents actually improve your day, and which are speculation wearing a product page. The honest answer is that most of the category is the second thing, and our own scoring shows it. Before the map, it helps to know what an agent is made of.

What an agent actually is

Here is the line the marketing skips. An agent takes a goal and acts on it over several steps, using tools to browse, call APIs, run code, or move funds, and it changes something beyond the text it prints. A chatbot answering one prompt fails that test. So does a single model call behind a ticker, and so does a character with a cartoon face that posts and trades on a fixed script.

An agent is a stack you assemble, and the layers split across ground we have already mapped. The layers it runs on are the Sovereignty Stack: the model and the inference that make up its brain, and the compute that hosts it. The way it pays is the Agent Commerce Stack: on-chain identity, a wallet, and standards like x402x402An open payment protocol from Coinbase that repurposes the long-dormant HTTP 402 status code. A server responds with a price, the client pays in stablecoins on-chain, and the request is fulfilled. No accounts, no API keys, no card details.Like a vending machine for HTTP. The endpoint says "pay 1 cent for this", the agent drops in a coin, and the document comes out. Settles on a blockchain underneath, but the payment layer is invisible to the caller.Read more → for settling work. This piece is about the layers in between, the ones that turn a model into an agent: the runtime, the tools, and the agents people ship on top.

The point of the layers is that sovereignty is decided at each one, not once. Own the runtime and you can fork it. Rent the brain and you can still switch providers. Rent every layer and your only recourse is to cancel the subscription.

I run Agent Zero, and it’s the whole argument in one setup. The runtime is mine, self-hosted in Docker on my own machine. The tools are mine, the skills and MCP servers I wire in. The host is my hardware.

The brain is whatever I point it at. Aim it at Venice or Morpheus and the model and inference are sovereign too; aim it at a frontier key and I trade that sovereignty for raw capability. No agent token appears anywhere in that assembly, and that is the tell for the rest of this page.

The bar a frontier agent already sets

Start with what you can use this afternoon without 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.Read more → or a token.

Anthropic, OpenAI, and Google have all shipped agent modes on their frontier models. They browse the web, operate a computer from screenshots, call tools, and chain multi-step tasks under supervision. Anthropic’s computer use and the equivalents from the other two labs are the most capable autonomous agents most people will touch this year. They are closed, hosted, and rented by subscription.

That sets the bar. Any specialised agentAI AgentAn AI system that takes a goal and acts on it across multiple steps, calling tools, APIs, or other models to get there, rather than just answering one prompt. In crypto, agents often hold a wallet and transact on-chain.Less like a calculator you punch a sum into, more like a contractor you hand a brief. You say what you want done; it plans the steps, picks the tools, and comes back with a result, occasionally making a mess if the brief was vague.Read more → has to clear the usefulness of a frontier agent you can already run, or offer something the closed model can’t: privacy, ownership, an open weight you control, or access to a market the labs won’t serve. The second question to ask is generality. A frontier agent does many jobs passably; a vertical specialist does one job, and either beats the generalist at it or has no reason to exist. Hold every project on this page against both lines.

The categories of agent, by the job they do

Sort agents by the job, not the marketing. Five jobs cover most of what ships today:

• Assistant and productivity. Research, drafting, inbox triage, scheduling, code. The frontier agent modes own this outright. No token product comes close on raw capability.
• Trading and on-chain (DeFAI). Autonomous agents that move money on-chain. Giza’s ARMA on Base, Olas prediction-market bots on Gnosis, Warden’s Agent Hub. This is the one vertical where a specialist agent has a clear reason to exist.
• Content and social. Posting bots and character agents. ElizaOS character files and the Virtuals launchpad cohort live here. Cheap to make, easy to flood, hard to make matter.
• Autonomous-economic. Agents that hold a wallet and pay other agents. The plumbing is shipping (x402, ERC-8004 identity, ERC-8183 escrow) but the demand isn’t. We covered the gap in The Agent Commerce Stack.
• Robotics. Embodied runtimes on physical hardware. Openmind’s OM1 runs on Unitree and UBTech robots. Early, capital-heavy, and a different game from the software agents above.

The pattern: the generalist wins the broad jobs, the specialist only earns its keep where it touches money or hardware the labs ignore.

The marketplaces, compared

Here is the agent-category cohort we have reviewed in full, against the questions that matter: what it actually hosts, whether you control the inference, and whether the usage holds up when you check it.

A few of these deserve a verdict in their own right, because they are the archetypes.

Virtuals is the launchpad archetype. The pitch was “pump.fun for AI agents,” and it delivered exactly that: trivially easy to mint a tokenised agent, a 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.Read more → for every one, and over eighteen thousand of them now sitting mostly inactive.

C

Quadrant

Centralised value

42

Freedom

/100

D

68

Returns

/100

C

Verdict · Returns over freedom

The tokenisation infrastructure is well built. The agents on top of it are mostly zero-revenue tokens, and the inference runs on the team's servers.

Strengths

• + Multi-chain launchpad with serious liquidity engineering
• + Buyback-and-burn and the Unicorn launch system are credible upgrades

Risks

• − 18,000+ agents, the vast majority inactive
• − Daily revenue collapsed 97% within two months of the 2025 peak
• − GAME inference is fully hosted; without it the agent is just a token

Olas is the counter-example, and the most interesting one. It has measurable on-chain agent activity, open-source credentials, and a working protocol. The economics underneath haven’t caught up.

B

Quadrant

Sovereignty play

62

Freedom

/100

C

35

Returns

/100

F

Verdict · Freedom over returns

Measurable on-chain agent usage and a properly open stack, undercut by a marketplace that turns over almost nothing and a token down 99.6% from its high.

Strengths

• + 14.5M on-chain transactions, 600+ daily active agents
• + Open Autonomy stack is self-hostable

Risks

• − Negligible lifetime marketplace turnover
• − 99.6% token drawdown from all-time high
• − Insider allocation with weak vesting

OpenServ is the hype archetype, and the clearest case of marketing running ahead of the product. The underlying research is sound. Almost everything wrapped around it overstates.

D

Quadrant

Avoid

24

Freedom

/100

F

34

Returns

/100

F

Verdict · Weak on both axes

A credible research paper surrounded by claims that don't survive a check: a 122x headline against a 74x paper, '400+ services' that resolve to about 31, and a flagship proof nobody can locate.

Strengths

• + The BRAID preprint is legitimate and has a named external co-author
• + The TypeScript SDK is properly open

Risks

• − No verifiable revenue; reasoning API still in private beta
• − Adoption story is roughly launchpad micro-caps dressed up
• − No security audit for a token live on two chains

The agents worth running don’t lead with a token

Run down the cohort above and a thing jumps out: the agents you would actually use aren’t in it. The useful ones increasingly ship as a feature inside projects we file under privacy or inference, where the agent is something you run and the token is a separate question. Three are worth naming.

Venice agentic chat is the clearest example, and the one most readers can use today. Venice made its agentic chat (Chat V2) the default chat experience in May 2026: a multi-step assistant that searches the web, calls tools, and generates media inside one conversation, running on private inference with no logged account. The same toolset ships as a Venice MCP server (31 tools) you can wire into your own runtime. Venice grades 5.7 on Freedom and 6.8 on Returns, above anything in the agent-token cohort, and this is the privacy agent the rest of the page keeps pointing at. Venice review.

Heurist Mesh is agent tooling rather than an agent, which is exactly why it’s useful. It’s a marketplace of 100-plus crypto-analytics skills (DexScreener, Etherscan, Arkham and the like) you plug into any agent over MCP, with USDC settled agent-to-agent through x402, live at mesh.heurist.ai. Good plumbing for the tools layer if you’re building. The HEU token is a separate and weak question, grading 3.8 and 2.5. Heurist review.

NEAR AI is the bet to watch rather than the agent to run yet. Its IronClaw runtime runs agents inside hardware enclaves (Intel TDX plus NVIDIA confidential compute) and ships alongside Private Chat in the near.com app, per NEAR AI. The agent capabilities are still thinner than the infrastructure underneath them, but the confidential-compute angle is the right one and NEAR grades 6.3 and 7.3. NEAR review.

The pattern is the tell. The best agent you can run today sits inside a privacy project and asks nothing of its token, while the launchpad cohort sells you the token and hopes you don’t check the agent.

Hype versus solid

Plot the agent-token cohort on our two scores and the verdict is hard to miss. Every agent-category project we have rated lands in the lower half on at least one axis. Not one reaches the “best of both” quadrant.

A Best of both 0

Empty.

B Sovereignty play 2

• Olas 6.2/3.5
• Openmind Review 5.6/3.7

C Centralised value 2

• Fetch.ai / ASI Alliance 5/5.5
• Virtuals Protocol 4.2/6.8

D Avoid 4

• ElizaOS 5.2/2.7
• Giza Review 4.2/4.4
• OpenServ 2.4/3.4
• Warden Protocol 4.7/4.8

That uniformity is the finding, and it explains why the section above matters: the agents worth running (Venice, NEAR) sit off this chart, in projects filed under a category other than “agent.” The cohort that brands itself as agents grades poorly, set against the Freedom and Returns methodologies.

Now separate what holds up from what doesn’t, using the same discipline we apply everywhere: a claim is independently checkable or it’s a press release.

Holds up when you check it:

• Olas: 14.5 million on-chain transactions and 600-plus daily active agents on Gnosis Chain (on-chain, as of June 2026).
• Giza: live ARMA fees recorded on DeFiLlama, a small base that has fallen sharply since launch (DeFiLlama, as of May 2026).
• Warden: roughly 35 to 37 million on-chain transactions confirmed by independent RPC sampling (as of April 2026).
• Venice: agentic chat shipped as the default chat in May 2026, with a published MCP server of 31 tools (Venice changelog, May 2026).
• Bittensor subnets: Chutes serves live inference cross-checkable through its OpenRouter listing; Templar published Covenant-72B, the largest decentralised pre-training run to date (March 2026).

Doesn’t survive checking:

• Virtuals: eighteen thousand agents, the vast majority inactive zero-revenue tokens.
• Warden: the 60-million “agent runs” headline folds in off-chain counts that can’t be verified, and the team admitted purging six to seven million bot users.
• OpenServ: the “400+ services” claim resolves to about 31 tools, the flagship proof is unlocatable, and the touted adoption is a cluster of launchpad micro-caps.
• Fetch / ASI: no public usage or revenue dashboards for any product after years of operation.

The pattern repeats across the cohort. The projects with the loudest user numbers tend to be the ones whose numbers thin out under a check. The quiet ones with on-chain activity you can sample yourself tend to be the ones doing the work.

What actually improves your life right now

Strip out the tokens and ask what you would actually run, sorted by what you are trying to get done:

• To think and make faster: a frontier agent mode, Claude, ChatGPT, or Gemini. The most useful agent for almost everyone, and it has no token.
• For a private assistant you control: Venice’s agentic chat. It does the web-searching, tool-using, multi-step work of a frontier agent on private inference with no account to suspend, and you can point it at your own stack through its MCP server. The closest thing to a frontier agent that doesn’t log you.
• For autonomous on-chain tasks: a specialist DeFAI agent, eyes open. Giza’s ARMA rebalances stablecoin positions across Base lending markets on its own, and its fees show up on DeFiLlama against a small, declining base. Treat it as a live experiment with your capital, not set-and-forget yield.
• For building your own: an open runtime plus open tools. ElizaOS or Agent Zero for the loop (self-hosted, you own it), Heurist Mesh for crypto skills over MCP. You self-host, you own it, the token is beside the point.
• For privacy in front of a plain model: a decentralised inference client like Morpheus Node Neo, which puts inference behind a self-custody client and plugs into tools like Cursor. Not an agent itself, the private pipe an agent or app runs over.

For most readers that is a frontier assistant, Venice’s agentic chat when the work is sensitive, and at most one specialist where you have a specific job. A wallet full of agent tokens doesn’t belong on that list. Buying the tokens is a portfolio bet on a category where nothing we have rated in it earns better than a C. The tools above are what change your week.

Why sovereignty changes the maths

If the closed frontier agents are the most capable, why bother with the decentralised stack at all. Three reasons, and none of them is the token.

The first is privacy. A hosted agent reads your prompts, your files, and your patterns. A self-custody client running decentralised inference, or Venice’s agentic chat on private inference, does the same job without a logged account behind it. For anyone whose work is sensitive, that is the difference between a tool and a liability.

The second is ownership. An open runtime or a permissionless network can’t raise your price overnight, change its terms, or switch you off. You hold the keys at the layers you control, and you can move. That option costs you nothing until the day you need it, and then it’s the only thing that matters.

The third is reach. The labs won’t serve every market, every jurisdiction, or every use they consider off-limits. A permissionless inference network doesn’t get to decide who you are. For a large slice of the world, that is the whole point of decentralised AI.

The construction PM’s take

I read this the way I read a build. An agent is a stack you assemble, the same way a building is a frame, services, and a fit-out from different trades, each from its own supplier. The frontier labs have poured the slab and framed the house. It works, you can move in today, and you are renting from a landlord who can change the locks.

The tokenised-agent marketplaces are spec houses thrown up fast to sell, with show-home photos that don’t match the rooms inside. Eighteen thousand of them on one estate, almost all empty. The launchpad sold the lots; nobody built much on them.

The handful worth your time do quiet, checkable work, and most of them aren’t filed under “agent” at all: a private agentic chat inside a privacy client, a DeFi agent with fees on a public dashboard, an on-chain agent protocol with transactions you can sample, an open runtime you host yourself. None of them lead with a token. All of them give you something a closed frontier agent can’t.

Fact: no agent-category project we have rated earns better than a C on either axis, and most grade D or F. The most capable agents available are closed frontier models with no token. The agents worth running on the decentralised side sit inside projects whose category isn’t “agent”, Venice’s agentic chat the clearest.

Take: assemble the stack rather than buy the badge. Use a frontier agent for raw capability, Venice’s agentic chat when the work is sensitive, an open runtime like Agent Zero if you want to own the loop, and one specialist only where it touches money or hardware. Ignore the launchpads. The agent that improves your life is the one you would still run if its token went to zero.