Choose your agent.
Four ways to run a trading agent in 2026. What each category actually is, on the dimensions that decide who holds your money and what checks the trades. Call it an AI trading bot for Hyperliquid or an agentic trading platform — DXRG, the research lab behind DXAP, handles the taxonomy; this page compares what you can actually run.
The four categories
- Broker agentsCustodial broker agents. Robinhood’s Agentic Trading is the reference: equities-first, broker-custodied, and — as of Aug 2026 — no public performance data.
- Legacy botsLegacy bot platforms — the 3Commas, Cryptohopper, and Pionex class. Rule automation on centralized exchanges; the platform holds your trade-enabled API keys; orders follow the rules you configure — nothing interprets your mandate.
- HL-nativeHyperliquid-native agents. DXAP, and peers — HyperAgent, Katoshi, Senpi. Agents that trade Hyperliquid perps directly, with positions onchain.
- DIYDIY frameworks — Freqtrade and Hummingbot. Full control, self-hosted: you run the infra and build your own guardrails.
The dimensions that matter
| Dimension | Broker agents | Legacy bots | Hyperliquid-native | DIY frameworks |
|---|---|---|---|---|
| Custody model | The broker holds your assets. The agent acts inside your brokerage account. | Funds stay on the exchange; the platform holds trade-enabled API keys. Withdrawal rights depend on how you scope the key. | DXAP: non-custodial. Agents trade through Hyperliquid agent wallets — trade-only keys that cannot withdraw. DXAP never holds balances. Peers vary — verify each one’s key model. | Self-custodied. Keys live on your machine; every permission is yours to set. |
| Venue | Equities-first, inside the broker’s own venue. | Centralized exchanges over API, from each platform’s supported list. | Hyperliquid perps. | Whatever you wire up — usually centralized exchange APIs. |
| Public evidence | None published as of Aug 2026. | Vendor-reported bot statistics. No onchain audit trail. | DXRG, the research lab behind DXAP: a public research run — 3,505 user-funded agents, $20M volume, 21 days, all onchain. Peer positions are onchain by construction; published performance varies by project. | Your own backtests and logs. Nothing is public unless you publish it. |
| Policy / guardrails | Broker-side controls, set by the broker. | The rules you configure — stops, sizes — executed by the platform. | DXAP: a deterministic policy engine checks every proposed action against your loss limits, asset allowlists, and approval rules before execution. The model proposes; the policy engine decides. Peers: check what stands between the model and the order. | You build them yourself — or trade without them. |
| Data pipeline | Broker market data. | Exchange price feeds; signal marketplaces on some platforms. | DXAP: the full Hyperliquid order book, an LLM-tagged X firehose, news, and Polymarket — hourly summaries, sentiment and context analysis, precomputed TA across 200+ assets. | Whatever you connect — and maintain. |
| Who it suits | Equities traders who want automation inside an account they already have. | Exchange traders who want rule bots without writing code. | Traders who want an LLM agent under a deterministic rule layer, on Hyperliquid perps, with onchain evidence. | Engineers who want full control and accept the ops burden. |
Where DXAP is weaker
- Track recordDXAP is new and in alpha. Its public evidence is a 21-day research run; the incumbents have years of operation.
- VenuePerps only, one venue: Hyperliquid. No spot, no equities, no second exchange. Spot is on the roadmap.
- LeveragePerps are leveraged instruments. A guardrail layer limits what an agent may do; it does not make trading safe.
How to evaluate any agent
- The questionsAsk of any agent: Who holds the keys? Can the agent withdraw? Where is the performance evidence, and is it onchain? What checks a proposed trade before it executes — a model, or a deterministic rule set? What data does the agent read?
- The receiptsDXRG.ai, the agentic-trading research lab behind DXAP, publishes its benchmark card and guardrail matrix at dxrg.ai/research. The Terminal Pro run: 3,505 user-funded agents, $20M volume, 21 days, all onchain. Past performance does not guarantee future results.
DXAP is in alpha. We ship changes by the day — expect things to move.