BlackRock just published a paper on LLM multi-agents for portfolio construction. The hedge fund of the future = a tiny team + a swarm of agents. AI-native trading is inevitable. That’s why we’re building ENVY for crypto. arxiv.org/abs/2508.11152

ENVY update: we’re going Solana-native (HyperLiquid stays). Sign in with your Solana wallet, deposit to a smart contract, and let your agents trade on Drift. Each agent can buy back its own Solana meme-coin tied to PnL. Bring your wallet, not your Python.

Look at ENVY Library of predictors for the main cryptocurrencies! Going live this week!

👤 We Analyzed 10,000 Hyperliquid Traders.
The Results Are Brutal.

- Only 16.5% of traders are in profit.
- 73.8% lost money.
- 27.2% lost more than 85% of their capital.
- The median result: –$67.

Outcomes closely track position size:
Small accounts lose systematically, while larger are more likely to stay profitable, reflecting structural changes in perpetual markets.

👉Check out the full analysis here.

ENVY Hackathon Overview 👤

- 100+ strategy predictors competed across 22 pairs.
- Each was stress-tested on 2–4 years of data across bull markets, bear markets & volatility spikes.
- Only 22 predictors passed the full evaluation.

Key finding:
- The predictors remained stable across all market regimes.

All 22 validated predictors are now live in the ENVY library, expanding the signal stack with new momentum, reversal, volatility & microstructure signals.

👉 Full review is available here.

👤 Can AI trade gold?

Agents aren’t limited to crypto.
The same architecture may work across equities, metals & commodities.

◾ Data plugs in as modules: prices, volume, volatility, macrosignals. Predictors read price action, regime shifts & news without changing the core logic.

◾ Execution and risk work the same across markets: position sizing, stops & drawdown limits are enforced system-wide.

The intelligence remains consistent, only the context changes.

👤 ENVY x ERC-4626 Architecture

Each agent runs on ERC-4626 - a tokenized vault model on Ethereum.
Capital is pooled inside the vault w/ PnL reflected directly in the share price.

It functions like an on-chain fund:
• Performance is positive → share value increases.
• Losses occur → share value declines proportionally.

◾️ Why ERC-4626?

- Composability - integrates with DeFi protocols that support the standard.
- Transparency - deposits, withdrawals & performance are auditable on-chain.
- Capital efficiency - users can access one strategy via share-based ownership.
- Standardization - proven structure that simplifies integration.

👤 Agents use a Frames & Chips model.

Frames define core logic - memory, forgetting speed, learning style.
Once launched, a Frame is fixed, giving the agent stable behavior.

Chips - modular analytics modules: price direction, volatility, regime detection, execution.
Chip versions don’t change after release, so results stay consistent.

▪️Chips are kept in a growing shared library, with new versions added as models improve.

Agents can switch to newer Chip versions without changing their Frame, and Frames define which versions an agent is allowed to run.

▪️This setup keeps the agents stable and lets them update safely.

👤Data proves the pattern:
most perp traders lose.

The problem is structural mistakes in leverage, risk management & emotional decision-making.

▪️Profitability comes from disciplined systems - without impulsive trading.

👤 Why LLMs can’t trade?

LLMs are strong at reasoning, but they ain’t built for trading:
no structured market context for order books, PnL, funding or risk.

Markets are combinatorial.
Edge comes from interactions between signals, not isolated ideas.
LLMs don’t explore this space, but generate strategies that sound right.

▪️Risk is a core gap.
Capital exposure & position sizing need strict control, even logical outputs can break basic trading constraints.

An LLM is just one module in a broader system, not an intelligent trading stack.