The market for crypto trading bots is, to put it kindly, mature in its problems. There's a familiar pattern: an account on Telegram, a screenshot of a recent winner, an invitation to deposit, and a black box that may or may not be running anything at all. Sometimes it's a real trader. Sometimes it's a spreadsheet. Often it's a fee-eating loop on a Binance API key. The user has no way to know.
Engine is built on the opposite assumption: that you should be able to see, in real time, every read, every rule check, every order, every fill the agent generates on your behalf, and that any product that won't show you those things is hiding something that's going to cost you money.
This essay is the long version of why.
Why opacity is profitable for the vendor
Start with the incentive. If you sell trading software, the cheapest version of the business is a fixed monthly fee with no transparency. You charge the user, you don't tell them anything, and they decide whether to renew based on their P&L screenshot. That's information they have to interpret in isolation, with no way to validate it against the strategy they thought they were buying.
This works for the vendor for three reasons.
You can hide losing months in the average. A monthly subscription means the user evaluates you on a rolling window. If you lose money in May and make money in June, the user remembers June. If you'd had to publish every trade in May, the user would have left in May.
You can charge for things that aren't happening. A "fully managed AI trading bot" can be a cron job that holds spot ETH. A "research-grade signal feed" can be a Twitter scraper. Without a decision log, the user has no way to tell whether the model is doing anything at all, much less anything complex.
Disputes are unfalsifiable. When a user asks why a trade lost money, "the model thought there was an edge" is an answer that ends the conversation. The user can't push on it. They can't verify it. They can either accept it or leave; there's no in-between.
These are good business reasons, and they're why most of the products in this market work this way. They are, however, not aligned with the user's interest. The user wants to know.
Why transparency is profitable for the user
The mirror argument: an opaque trading product cannot, structurally, be improved by its user.
Trading is a long feedback loop. You change something, you wait, you see what happens, you update your model. If the only thing the user sees is the P&L number, the loop is too coarse to learn from. They can tell whether the strategy is making or losing money, but they can't tell why, and so they can't tell which part to change.
Give the user the decision log and the loop tightens dramatically. They can see that the strategy is overconfident in low-vol regimes, or that it's getting filled at bad prices in thin perps, or that one specific rule is producing 80% of the trades and 20% of the PnL. With that information they can edit the file, restart, and see whether the change worked. Without it they have one knob, which is "use the bot or don't."
That's the case in one sentence: transparency is what makes a trading product editable, and an editable product is what lets the user actually get better at trading.
What "transparent" means when there's still a model
We're not the first product to claim transparency. The word gets diluted, so let me say what we mean by it concretely.
Three things are non-negotiable in Engine.
-
The decision log shows every read, every check, every order, every fill. Not summaries. Not "agent took 3 trades today." The actual sequence: at 09:14:18 the agent saw funding at -0.012% on ETH-PERP, matched rule R3, sized 2.4% NAV, set the stop here, took the fill there, paid this much in fees. Replayable, scrollable, exportable.
-
The playbook is in plain English in a file you wrote (or forked). When the agent acts, it explains the thesis, evidence, and invalidation it used. You can open the file, read the playbook, understand why the trade happened, and edit the guidance if you don't like the answer. The model isn't inventing a trader from scratch; it is applying the trader you described.
-
Funds stay in your Hyperliquid vault. The agent has a trade-only signing scope: it can submit orders, it can't withdraw. You retain the right to pull the rug on Engine at any time, with one click, and your funds are still where they were before you ever logged in.
Those three are the floor. We won't ship a feature that compromises any of them.
Three things are not transparent, and that's fine.
-
The model weights are not transparent. We are not going to publish the parameter file of the model that interprets your strategy and reads market data. This would be both useless to you and adversarial-vulnerable in production. What you get instead is the trace: the agent's actual sequence of actions, with the rules cited.
-
The exact prompt graph is not transparent. The model has internal scaffolding (sub-prompts for risk, for routing, for explanation generation) and we're going to keep iterating on those without notifying you of every change. What stays stable is the contract: your file is the source of truth for what the agent does; the model is the engine for how.
-
The set of features the agent is reading is partially transparent. The agent reads more than is in your strategy file; it has to, to make sense of what your file is saying. (When you say "skip if BTC realized vol > 80%," the agent has to know how to compute realized vol.) We document the full feature set, but we don't promise it's frozen.
The point of these distinctions: transparency isn't about making everything legible. It's about making the parts that affect your money legible. The trace, the rules, the custody. Everything else is implementation, and implementation should be allowed to evolve.
The non-custodial part
A specific note on custody, because it's where most of the bot industry's worst stories live.
When you use Engine, your funds never leave your Hyperliquid vault. Engine doesn't have a hot wallet pool. Engine doesn't have access to your private keys. Engine doesn't have a "deposit address" on our side. What we have is a trade-only EIP-712 signing authority that you grant when you deploy a strategy, that we use only to submit orders, and that you can revoke at any time directly at the venue without our cooperation.
This means the worst-case scenario for an Engine user (every Engine user, simultaneously) is that they revoke their signing authorities, withdraw from Hyperliquid, and forget we exist. Their funds are intact. We'd lose the business; they wouldn't lose the money.
This is the right design for an industry that has, repeatedly, demonstrated what happens when a counterparty holds your funds and that counterparty's incentives diverge from yours. We'd rather be the kind of product you can stop using.
The boring conclusion
We don't think transparency is a marketing position. We think it's the only structural answer to the problem of how an outside party can credibly trade on your behalf without slowly stealing from you. The decision log makes the strategy editable. The plain-English rules make the agent debuggable. The non-custodial scope makes the relationship reversible.
Black-box AI trading is going to keep being sold, and it's going to keep working as a business for the people selling it. We're betting that a meaningful number of users would rather have something that they can read, edit, and walk away from, even if it costs slightly more to build that way. So far, the bet is going fine.
If you want to try the editable-and-walkable-away version, the agent is here. If you want to keep an eye on what we ship around transparency specifically, this is the section it'll show up in.