Ask ten people in crypto what an agent trading platform is and you will get three answers.
Some mean a leaderboard where language models compete with real money. Some mean a vault that routes stablecoins between lending markets while you sleep. Some mean a launchpad where anyone can mint a token with a chatbot attached and call it an agent.
Those are separate businesses with separate risks, and the word "agent" is doing a lot of work to hide the differences. The useful comparison is not which one advertises the best returns. It is which one can be checked, and who holds the keys while it runs.
The Reality Check Nobody Wants To Frame
Start with the most-quoted experiment in the category. Nof1's Alpha Arena gave six frontier models $10,000 each in real capital and set them loose on crypto perpetual contracts through Hyperliquid, with identical prompts and data. Over roughly two weeks, Alibaba's Qwen3 Max finished up 22.3% and DeepSeek Chat V3.1 up 4.89%. Everything else went the other way. Claude Sonnet 4.5 lost 30.81%, Grok 4 lost 45.3%, Gemini 2.5 Pro lost 56.71%, and GPT-5 lost 62.66%.
The behavior underneath matters more than the ranking. Win rates clustered between 25% and 30% for every model, Gemini placed 238 trades against Claude's 38, and fees ate a serious share of the outcome, with Qwen paying $1,654 in costs even as it won. Nof1's Jay Azhang later said the setup was deliberately unkind, feeding models numerical time series through limited context windows.
Two weeks is one path through one market, so it proves nothing about durable skill. What it shows is that the gap between models is enormous, and that a platform quoting one agent's recent return is telling you almost nothing.
Arenas Have Multiplied, With A Trade-Off
The benchmark corner has grown since. TradeRank runs rolling seasons with 16 frontier models, and its ninth season opened on Sep. 12, with each handling $10,000 across ten major cryptocurrencies and fifty large-cap US equities, with every trade and every line of reasoning published.
That is a far better sample than a fortnight. The catch is that the capital is simulated, so nothing there pays real slippage or has to exit a position that has stopped being liquid. Alpha Arena's numbers hurt more because they were real. TradeRank's cover more ground because they are not. Any platform citing arena performance should say which kind it means.
Where The Money Actually Sat
The most honest case study in the category is a project that closed. Giza's ARMA agent automated stablecoin yield across Aave, Morpho, Compound and Moonwell, and it grew to around $4.85 million in managed capital. In February 2026 Giza announced ARMA would be retired, gave users a month to move, and shut it down on 26 March with funds returned to wallets and no exit fees. Its successor agent held roughly $14,000 in on-chain balances by June.
That cuts both ways. The wind-down was handled properly, which is more than many projects manage. It also shows how thin real demand was. A category that talks in billions had its best-known consumer agent peak below $5 million.
The Question That Separates Them
Once you stop comparing returns, custody becomes the axis that matters. Most consumer-facing agents now run non-custodially, so funds stay in your wallet or your own smart account rather than on the platform's balance sheet. Neyro, HyperAgent, Kora, Almanak and Polystrat all work this way.
The finer distinction is what permissions the agent holds while it operates. HyperAgent runs on trade-only API keys with no withdrawal access, so the worst case is bad trading rather than a drained wallet. That is a real difference from an agent holding broad approvals on your tokens. Numerai sits outside the pattern, because it is a hedge fund with institutional capital and a public tournament, so retail players contribute models rather than trade their own money.
Also Read: What Serious DeFi Users Know About Oracles That Beginners Always Miss
The question is not what the agent can do. It is what it is allowed to do while doing it.
Nobody Compares The Place The Agent Trades
There is a fourth category that gets left out of these comparisons, and it decides more than the model choice does. An agent has to execute somewhere, and that venue sets both the speed of the fill and the identity of whoever holds the funds.
A centralized exchange gives an agent in-memory matching with no block to wait for, plus deep books and spot, futures and options in one place. The cost is that the exchange holds the money, you clear an identity check before trading at all, and the withdrawal rights on an API key vary by how it was issued.
A decentralized venue inverts that. Funds stay in your wallet, agent keys can be scoped so they cannot withdraw, and there is no sign-up beyond connecting a wallet. The cost is that every order waits on the chain, roughly 0.2 to 1.1 seconds to commit on the larger perp venues, and your order flow is public.
For a strategy trading once a day, that gap is noise. For anything reacting to a move it is the whole game, and it is why many agent products quietly run on centralised venues while describing themselves as on-chain.
Yellow Pro, built by Yellow, is one attempt to stop treating that as a choice. The architecture underneath it is a clearing network using ERC-7824 state channels, in which participants post collateral to smart contracts, open channels with counterparties, then exchange liability updates off-chain and commit only net settlement to the base chain.
The Block has described the same design as handling billions of off-chain messages a day at throughput comparable to a centralized exchange.
In practice, Yellow says that means self-custody through your own wallet, an agent working inside your own sub-account with no withdrawal rights, execution latency under a millisecond, spot and perps through a single connector, and sign-up by wallet or Google account with no identity check.
It also puts the agent connection at under five minutes, against the exchange route of clearing KYC, generating an API key and then wiring up a connector.
Those product figures come from Yellow rather than from a third party, which is the same caveat this article applies to everyone else. The state channel architecture is documented and independently covered. The specific latency and onboarding numbers are the company's own, and the sensible thing is to test them with a small allocation before handing an agent anything that matters.
The Token Layer Is A Different Bet
Then there is the part of the sector that trades as a narrative. Virtuals Protocol, the launchpad that defined the first agent mania, is down roughly 90% from its January 2025 peak, with daily revenue falling from about $1.02 million to under $500 by early April 2026. FET is off 91%, and combined agent token market value sits near $2.92 billion against more than $10 billion in late 2025.
Usage went the other way. Daily active on-chain agents passed 250,000 in early 2026, up around 400% year on year, and Microsoft reckons 80% of Fortune 500 companies now run agents. Adoption rose while the tokens collapsed. Owning an agent token is a bet on a token, not on the technology working.
How To Compare Them Properly
Four questions will sort most of the field. What permissions does it hold over your funds, and can it withdraw. Is there a verifiable record, meaning on-chain history or published reasoning rather than screenshots. How long is the track record, and did it run through a genuinely bad market rather than a quiet quarter. And what happens to your money if the team walks away, since ARMA proved that question has a good answer and a bad one.
None of that tells you whether an agent will make money. It tells you whether you can find out.
Read Next: Why Zero-Knowledge Proofs Could Become Crypto’s Most Important Technology
