AI trading agents · Hyperliquid · Pre-MVP
Describe your edge. The agent makes the calls — and the trades.
Automated trading is usually human-in-the-loop: you still make the discretionary calls — read the regime, decide whether to stand aside, time the entry — and the bot just executes the mechanical part. Edgena is agent-in-the-loop. You describe your edge once, in plain language; the agent applies the judgment every bar and executes automatically — inside risk caps the execution engine enforces in code before any order reaches the exchange.
Agent-in-the-loop, not human-in-the-loop
Rules are the easy half. Judgment is the edge.
Professionals run systematic rules inside a discretionary outer layer: rules govern entries and exits — judgment governs whether to be in the market at all. The rules you can automate. The judgment — is this the right regime? is volatility sane? should I stand aside today?— is the part most tools leave to you. For a part-time trader that layer is fatigue, emotion, and inconsistency. It's where strategies quietly die.
Edgena isthat judgment layer. You bring the edge; the agent runs the overlay every tick, without fatigue. Whatever your experience level, the agent's discipline at runtime is the same — your tier only changes how much Edgena explains while you build, never how it trades.
The systematic core
A thesis or a set of entry/exit rules — the edge you want to capture. Describe it in plain language; even a one-line thesis works. This is the part that can be written down and backtested.
The discretionary overlay
Every bar (5m or longer — not real-time): read the regime, weigh the event calendar and funding, honour your kill-conditions, and stand aside when conditions don't fit — then apply your rules only when they do. The judgment a pro applies by hand, applied consistently while you're away.
The hard floor
Every proposed order is validated against your risk caps before it can reach the exchange. Leverage, exposure, order rate, loss + drawdown halts — non-negotiable, enforced in code, identical in sim and live.
Edgena decides at bar cadence (5 min or longer), reading each market top-down across timeframes — not tick-by-tick. Each Skill commits to one trading style — day, swing, or position (below) — chosen up front; the one thing Edgena can't do is sub-minute scalping. Between decisions a position is protected by resting stops + your risk caps, not by Edgena watching. Discipline is not a profit guarantee: markets are adversarial and a backtest only describes the past. Edgena raises the floor on process, not the ceiling on returns.
Pick a horizon — and commit to it
Every Skill is one trading style. Chosen up front. Fixed for life.
Day, swing, or position — set when you create the Skill. It fixes the bars the agent decides on and how far up the timeframe ladder it reads for context. You choose the horizon; the platform never guesses it from an interval.
Day trading
In and out within the day. Fast intraday reads, flat-ish overnight.
Swing trading
Hold across many bars to capture a multi-day move; tolerate the noise inside it.
Position trading
Hold through short-term noise for the major trend. Decide rarely, deliberately.
And it's immutable. The fastest way a part-time trader blows up is switching styles at the first sign of trouble — abandoning a swing system mid-drawdown to day-trade, then dropping that the first red session. So a saved Skill's style is locked: stay loyal and let it compound, or start a fresh Skill for a different horizon. Discipline, made structural.
How it works
Talk to Edgena
Open the chat. Tell it what you want to trade. It asks 3–7 tier-appropriate questions, drafts a Skill in a preview pane, and runs a senior-PM-style critique before you save.
Backtest it
One click. Replay your Skill tick-by-tick against historical bars + news + funding. Get Sharpe, Sortino, win rate, profit factor, max drawdown — the standard quant metrics, computed daily-series.
Deploy on Hyperliquid
Promote to live. Each Skill gets its own Fly machine on its own schedule. The chat agent becomes your coach: ask why a position opened, what's coming that could hurt it, when to flatten. Confirm-card actions only — the bot never places an order itself.
Why Edgena
The engine is the safety promise
AI agents never call the exchange directly. They emit JSON proposed actions; the execution engine validates each against your Skill-defined risk caps (max position, max leverage, daily loss halt, rate limits) and rejects or executes. Same code in sim and live — sim is a faithful preview, not an approximation.
One chat, every job
Authoring, coaching, and ops all happen in the same conversation surface. Open it on a deployment page, the agent already knows what is running. Ask it to redeploy or start a backtest, it prepares a confirm-card; you click. The LLM stays out of the command path by construction.
It remembers what works
Every closed trade lands in a structured ledger with PnL, MFE/MAE, regime tag, and the reason the agent gave. After every Nth trade, a reflection job distills lessons. Your strategy stops repeating its own mistakes — without you having to babysit it.
Versioned, auditable, yours
Skills are immutable versions. Conversations persist. Decisions snapshot every tick: context the agent saw, tools it called, why it proposed what it did, what the engine said back. Trade your own funds via your own keys; Edgena is software, not a broker.
Stack
- Models
- Any LLM via OpenRouter — Sonnet 4.6 default; per-Skill swap; BYO API key.
- Exchange
- Hyperliquid perpetuals (mainnet). Cross-margin paper broker for validation.
- Runtime
- Vercel for chat + control plane. Fly.io for the long-running per-Skill trading process.
- Data
- Supabase Postgres + pgvector. Bars + funding + news + macro events; news-derived event calendar.
- Tools
- Built-in tool registry. MCP-compatible (curated catalog in MVP).
- Open
- Architecture, ADRs, and engineering docs published. Marketplace + multi-exchange post-MVP.