TL;DR

  • “Onchain AI agents” became the dominant crypto narrative in 2025 and has cooled meaningfully in 2026 as the picture has gotten clearer.
  • The honest taxonomy has three buckets: agents that hold wallets and trade, agents that automate DeFi operations, and agents that exist primarily as tokens with a chatbot attached. Only the first two are doing real work.
  • Real revenue is concentrated in agent-driven DeFi automation, MEV strategies executed by agents, and onchain payment rails for AI services. Most of the rest is meme economics dressed in technical clothing.
  • The structural question - “do AI agents need crypto rails at all” - has become a genuinely live debate. The answer in 2026 is “yes, but only for a narrow set of jobs, and most of those jobs are not what was being pitched.”
  • If you are evaluating an onchain AI agent project, the test is brutally simple: strip away the token and ask whether the agent does something useful. If the answer is no, the project is a token with extra steps.

How We Got Here

The phrase “onchain AI agent” started showing up in crypto Twitter in late 2024 and exploded in early 2025. By the middle of last year there were thousands of agent tokens, dozens of agent platforms, and a handful of agents with billion-dollar implied market caps doing things that would have embarrassed a 2010-era chatbot.

The pitch was seductive: a language model gets a wallet, a personality, and the ability to take actions onchain. It posts on social media, manages a portfolio, runs a DAO, makes deals with other agents, and earns its keep. Crypto provides the rails - identity, payment, ownership, automation - that the AI side of the world cannot offer natively.

Some of that pitch was right. Some of it was the loudest grift of 2025. The job of this post is to separate them.

The Three Buckets

Most of the projects calling themselves onchain AI agents fall into one of three categories. The categories matter because the economics are completely different.

Bucket 1 - Agents That Hold Wallets And Trade

This is the genuinely interesting category. An agent has a wallet, a strategy, and access to onchain markets. It executes trades, manages positions, and reports performance. The model behind it is real (usually a hosted frontier model, sometimes a smaller open-weights model) and the actions it takes are real onchain transactions.

Examples include trading agents on platforms like Virtuals Protocol, portfolio managers built on agent frameworks, and a long tail of bespoke trading bots that present as agents.

What works: small-scale, narrow-strategy agents trading on liquid markets. They work the way a well-tuned algorithmic trading bot has always worked, with the language model handling the decision logic that used to live in handwritten rules.

What does not work: agents pitched as autonomous portfolio managers that “learn” from market conditions and “evolve” strategies. The empirical track record of these is unflattering. They underperform passive benchmarks. They get rugged by adversarial market participants who know how to bait language models. The novelty is real. The edge is not.

Bucket 2 - Agents That Automate DeFi Operations

This is the boring, useful category and the one quietly making money. An agent monitors positions, manages collateral, harvests rewards, rebalances portfolios, and routes transactions across DeFi protocols. The “AI” part is mostly the orchestration layer - parsing intents, summarising state, choosing among options.

Examples include intent-based DeFi platforms, automated yield aggregators with agent frontends, and the new generation of “DeFi copilot” products.

What works: automation of repetitive, error-prone DeFi tasks. Approving the right token, checking the right price oracle, splitting a transaction across the right venues. These are jobs where a competent agent saves real time and avoids real mistakes. Users pay for them.

What does not work: automation pitched as autonomous strategy generation. Choosing which yield to chase is a judgement call about risk; the agent does not have better risk intuition than a competent human, it just has faster fingers.

Bucket 3 - Agents That Are Tokens With A Chatbot Attached

The largest category by count. An agent is launched with a token, a personality, and a Twitter account. The token has a market cap. The agent posts. People speculate. The “agent” never does anything onchain other than transact in its own token.

There is no useful work being done here. The economics are pure speculation. When the music stops, the token goes to zero and the bot keeps tweeting into the void.

This is most of the market by token count and a meaningful fraction of the market by liquidity. It is also where most of the noise comes from.

Where The Money Actually Flows

If you trace dollar flows rather than narrative attention, the picture inverts.

Agent-driven MEV. Agents that participate in onchain auctions, liquidate undercollateralised positions, and arbitrage between venues are quietly profitable. This is mostly invisible because it does not have a marketing budget. The teams running these agents do not need to advertise.

Automation services. Subscriptions to DeFi automation platforms, AI-augmented trading dashboards, and intent-based execution products. Recurring revenue, real users, modest but durable.

Agent-to-agent payments rails. This is the category that almost nobody is talking about and the one that has the strongest long-term thesis. When agents need to pay for compute, data, or services from other agents, traditional payment rails do not work - they require human accounts, KYC flows, and banking relationships. Stablecoins on fast settlement chains do work. The volume here is small but the growth curve is the most consistent.

Token launchpads and agent platforms. The platforms hosting agent launches make money on fees regardless of whether the agents work. Picks-and-shovels economics. Whether this is a good thing depends on your priors.

Agent token speculation. Massive in dollar terms during 2025, much smaller in 2026, still nonzero. Mostly redistributive rather than productive. People making money from other people losing it.

The first three are real businesses. The fourth is opportunistic. The fifth is a casino.

The Structural Question - Do AI Agents Actually Need Crypto?

This is the question the space has been carefully avoiding and that has become harder to avoid in 2026.

The argument for crypto rails goes like this:

  • Agents need identity that is not tied to a human KYC account.
  • Agents need payment rails that work without banking relationships.
  • Agents need state and ownership primitives that survive across services.
  • Agents need composability with other agents and onchain services.
  • Crypto provides all of these natively.

The argument against goes like this:

  • Most agents in production today are owned by humans or companies that already have identity and payment infrastructure.
  • Stablecoins are useful, but most “onchain AI” projects are not actually using stablecoins for anything important.
  • The composability argument matters mainly for agents talking to other agents - and most agents today are not talking to other agents.
  • The crypto-native parts of the architecture often add friction without adding value.

The honest answer in 2026 is both arguments are partly right. There are jobs where crypto rails are genuinely better - cross-border agent payments, onchain reputation for agents, programmatic access to liquidity, autonomous economic activity that does not have a human counterparty. There are also a lot of projects pretending to need crypto when they really just want a token.

The test is whether removing the crypto layer would degrade the product. For most agent-token projects, removing the token would actually improve the product. For a smaller set of genuinely useful agent infrastructure projects, removing crypto would break them. That second set is where the durable value is.

What To Watch In 2026

A few signals separate the projects with a future from the projects living off attention.

Recurring revenue. Does the project have users paying for something other than tokens? If yes, the floor under the project is real. If no, the floor is wherever sentiment puts it.

Agent autonomy boundaries. Is the agent making decisions that a human would otherwise make, or is it dressing up a human-driven workflow in agent language? The former is a product. The latter is marketing.

Use of stablecoins. Real agent economic activity tends to settle in stablecoins. Heavy reliance on the project’s own token for operational activity is a tell.

Talent profile. Who is on the team? Crypto-native traders pivoting to AI tend to ship faster than AI engineers pivoting to crypto. Both are present in the market. The blended teams - serious AI engineers paired with serious crypto engineers - are the most likely to produce something durable.

Honest disclosure of model behind the agent. Most “agents” are thin wrappers over hosted frontier models. That is fine. Projects that obscure this and pretend to have proprietary agent intelligence are usually overstating what they have.

What I Would Bet On

If I had to put the categories in priority order for where productive activity will happen over the next 12-24 months:

1. Agent-to-agent payment rails. Stablecoin-denominated, fast, settled onchain. Whether this layer is dominated by general-purpose chains or specialised agent-payment chains is still open. The category is real.

2. DeFi automation with AI-native interfaces. Intent-based execution, agent-managed yield, AI-augmented portfolio management. Quiet, useful, growing.

3. Onchain reputation and identity for agents. When agents need to prove they are who they say they are across services, onchain identity primitives become the natural answer. Early but coherent.

4. Agent marketplaces and discovery. As the population of agents grows, finding the right one and paying for its services becomes a real problem. Whether this gets solved by A2A-style protocols or by onchain marketplaces is unclear, but the problem is real.

5. Probably not - autonomous agent DAOs running businesses. This was the most ambitious pitch of 2025 and has consistently underdelivered. The technology is not the problem; the governance and accountability gaps are.

6. Definitely not - tokens whose only utility is being held by an agent. Time will sort these out.

The Honest Verdict

Onchain AI agents are real and small. The narrative made them sound real and large.

The disappointment over the last year has been a healthy correction. The projects that survived the cooling are the ones with actual revenue, actual users, and actual technical substance. The ones that did not survive were mostly tokens with personalities, and the world is not worse off for them being gone.

What remains is genuinely interesting. The intersection of AI agents and crypto rails is one of the few places where two technologies that get a lot of solo hype actually have a coherent story together. Most of the value will come from the boring versions of that story - payments, automation, identity - rather than the flashy ones.

If you want to understand where this space is heading, ignore the agent leaderboards and watch the stablecoin payment volumes between agent wallets. The first metric is theatre. The second is the economy.