conco adaptive bot.
A bot that operates Hyperliquid on mainnet, 24/7, without me touching it. The interesting part is not that it trades — it is how it learns while doing it.
The bot watches a small group of Hyperliquid markets on short intervals. On every candle it classifies the symbol's context — clean trend, sideways range, early momentum or a zone where the right move is to do nothing — and, based on that context, decides whether to open a trade, let it pass or close a position already in motion.
Everything runs autonomously: I don't approve trades, move stops or touch sizing. The system runs 24/7 with its own supervisor, reconciles itself with the exchange on every tick and protects itself if something doesn't add up. Human intervention is reserved for changing the criteria, not for operations.
The bot's core idea is that every operation, beyond being executed, should leave a trace. When a position closes — well or badly — the system archives it classified by context: what symbol, what market regime, what relative-strength zone it entered in, what volatility level surrounded it.
Over time, that archive stops being a passive history and becomes something active: a context-aware memory. Once a specific situation has enough evidence, the bot starts treating it differently. If a setup keeps getting stopped out by noise, it widens the stops. If a combination strings losses, it benches it for a while. If another performs consistently, it works it with more size.
The point isn't the rules themselves: it's that those rules are decided by the bot, out of its own operation. I don't pre-set the strategy — it emerges from what the market keeps teaching it.
Before chasing returns, the system tries to not lose capital. There are several chained protection layers and all of them are non-negotiable: per-trade size limits, aggregate exposure cap, a circuit breaker on loss streaks, daily and weekly drawdown guards, and automatic deactivation when behavior drifts out of its expected envelope.
These layers exist because the learning phase is, by design, conservative. I prefer a bot that starts slow and goes far over one that looks brilliant for a week and self-destructs the next. That is the principle the rest is built on.
I don't sell the code or the bot's recipe. But I am happy to talk about how it is designed, what problems I solved along the way and where a similar idea can be applied.
If you are building in this direction, if you have a use case that fits, or simply want to compare notes, reach out.