RESEARCH · HOW WE BUILD WHAT THE AGENT SEES
This is not market commentary. It is the unglamorous work of deciding what the agent is allowed to see: which numbers are trustworthy, how old is too old, what should stop a trade outright, and what gets written down afterwards. Cleaner inputs make for a more selective agent — and a decision you can actually read back.
Six inputs, each with a job: how price is moving, what the crowd is holding, where the big money is going, what the news is saying, which windows to sit out, and a record of what happened. Together they are everything the agent gets to look at.
We only ever read candles that have finished. Three views of the same market — a short one for timing, a medium one for the setup, a long one for the trend — so the agent is never reacting to half a candle that could still turn around.
Funding, open interest and the order book tell us what everyone else is already positioned for. That is context around the price picture, never a reason to trade on its own — crowded does not mean wrong, and expensive to hold does not mean about to turn.
When large holders move money into stablecoins they are stepping back from risk; when they move out they are leaning in. We read that as a slow gauge of mood, not a trade trigger. If the data is missing or stale it counts as neutral — it never pushes the agent into a position.
Headlines are stored whole, so any decision can be traced back to what was known at the time. They are used twice: as context before the agent enters, and as a warning while a trade is open. We care about catching genuinely bad news, not scoring every headline.
Scheduled events like rate decisions are treated as stop signs, not suggestions. The agent stands down around them rather than guessing through a spike it has no edge on — and it does not even ask the model, so nothing is spent while it waits.
Every cycle stores what the agent saw, what it decided, and what the market did next. That record is what makes a decision reviewable long after the fact — you can always go back and see exactly what it knew at the time.
There is no shortage of crypto data — most of it is noise. The work is deciding which inputs are allowed to stop a trade, which only get a whisper, and what the agent should do when one of them goes quiet.
Continuous trading hours, dense funding/OI history, and usable on-chain risk-appetite lenses make BTC the highest-signal environment for agent evaluation. Secondary assets are admitted only when they meet the same evidence-completeness bar.
When big holders rotate into stablecoins they are getting cautious; when they rotate out they are getting confident. We use that as a slow read on mood. It can make the agent more careful, but it can never talk it into a trade the price picture does not support.
Some inputs are allowed to stop the agent; others can only shade its view. If a source goes down we mark it unavailable and carry on — we never fill the gap with an average, because a made-up number is worse than a missing one.
News is noisy, so a headline on its own is never enough. While a trade is open the agent re-reads the stored news alongside what price is actually doing, and it can only exit early when both agree something has genuinely gone wrong. Being too twitchy costs you a good trade; being too slow costs you money.
An agent can only be as good as what it can see. BTC trades around the clock with deep books and rich derivatives data, so that is where the evidence is most complete. New markets get added when they clear the same bar — not to make the list look longer.
A general model brings the reasoning. Alongside it we train our own on the history this platform generates. Either way the job is the same: the right evidence in, firm limits around it — a careful agent rather than a confident one.
The agent can only be as good as what is put in front of it. Most of the work is unglamorous: keeping inputs clean, deciding what counts as too stale to use, and designing the rules that stop a trade before the model is even asked. Better inputs mean fewer forced trades and clearer post-mortems.
Not the biggest possible return. We want it to be right when it commits: a high bar for taking a trade, an honest record of the ones it passed on, and reasons you can read afterwards. A quiet week where it correctly sat out is a good week.
A general model reasons well but knows nothing about how this system has actually behaved. Ours is trained on exactly that — every decision, the evidence behind it, and what the market did next — using real candle outcomes rather than hand-labelled examples. It is retrained weekly and scored on data it has never seen.
Leverage, risk per trade and the stand-down windows are not learned from results. They are set in code and they stay there. Research can make the agent more careful — it is never allowed to make it bolder by quietly drifting a number.
Notes on how this is built and how we check it. Educational — none of it is trading advice.
How we line up price, derivatives, capital flow and news into a single snapshot — and mark plainly what is missing, so the agent is never handed a view that looks more complete than it really is.
Why money moving in and out of stablecoins gets a say in caution but never a vote for entry, how we handle data that has gone stale, and what broke when the provider degraded.
Most candles end in a decision not to trade. We store the reason and the evidence behind every one of them, so a pass can be reviewed as carefully as a position.
Where we set the bar for an early exit: stored news read together with what price is doing, with a deliberately strict threshold so a frightening headline alone cannot close a healthy trade.
The case for starting where the data is richest. Continuous trading, deep books and usable capital-flow data make BTC the best place to do this work until other markets clear the same bar.
▸ Research materials describe system design and evaluation. They are not investment advice. Crypto is volatile; losses are possible. See risk disclosure.
See how six inputs become a single long, short or skip call — and how the trade is watched once it is open.