Risk Management
Risk is limited in five layers, from the position rules you set on /create to bounds the vault contracts enforce on every report, and each layer works even if the one above it fails.
An agent's risk controls are not one setting. They stack: the rules you choose, the limits built into each strategy engine, checks that can veto a buy, validation of the rules themselves, and caps in the on-chain contracts. This page walks through each layer, then compares the main risks of every template.
Layer 1: position rules (token-filter agents)
Uniswap V4 launches and Build your own agents hold one token at a time under six rules. You set them on /create; the bounds are enforced in code.
| Rule | Range | Effect |
|---|---|---|
| Position size | 5% to 50% | Caps what one token can cost the vault: at 10%, a total wipeout of the token is a 10% vault loss |
| Take profit | +5% to +500% | Locks in a gain once reached |
| Stop loss | -90% to -2% | Exits once the position is down this much |
| Hold while in top | 0 to 10 | Keeps a winner while it still ranks; lower values rotate sooner |
| Min hold | 0 to 240 min | Stops costly churn; never delays take profit or stop loss |
| Max round-trip cost | 1% to 30% | Refuses tokens whose fees, creator tax, price impact and gas, in and out, exceed this at your position size |
How they interact:
- Exits are checked every 30 seconds against the live price. The exit is booked at the modeled sell price, after fees, tax and price impact, so a stop at -30% usually realizes somewhat worse.
- Stop loss is not a guaranteed price. A token that halves between two checks exits at the next check's price.
- Max round-trip cost scales with size. Price impact grows with the position, so raising position size can quietly push tokens over your cost limit and out of the candidate list.
Layer 2: engine limits
Engine strategies have their own risk settings (full tables on Prebuilt Strategies):
- Concentration caps: Stablecoin yield's max per pool (20% to 100%) and pools at once (1 to 5); funding carry's markets at once (1 to 5); momentum's assets held (1 to 4).
- Exposure caps: DCA's max invested (10% to 100%) and tranche size (1% to 20%).
- Leverage: Perp trend allows 1x to 3x, with a stop loss of 2% to 20% of price that keeps the market closed until the signal flips.
- Trend filters: DCA, momentum and basket can step aside to stables when price falls below a moving average.
- Crowding and news guards: Perp trend can refuse a side whose funding costs more than a set APR, or that goes against strongly one-sided news (News sentiment tool).
- Cost awareness: Stablecoin yield skips a pool whose yield would not cover twice its gas; LP, yield and carry only move when a replacement pays a set amount more.
Layer 3: pre-trade checks
Guard tools run just before a token buy and can only veto it. With the Security scan, GoPlus data vetoes honeypots, tokens that cannot be fully sold, and buy or sell taxes over 10%. With the AI pre-trade check, Claude reviews the token and vetoes clear scams, copycats and dead narratives. Up to three picks are checked in order; a vetoed pick is skipped, not retried. If a check cannot run, the buy goes ahead, so guards reduce risk but never guarantee it. Details are on Security & Pre-Trade Checks.
Layer 4: rule validation
- Every value is checked against its bounds; out-of-range values are rejected, not rounded.
- Unknown keys, unknown parameters and unknown tools are errors.
- A token filter allows at most 24 conditions, and the whole rule set at most 4,096 bytes.
- If stored rules ever fail validation, the engine keeps trading on the last valid set.
- Rule changes after launch (a fix from a loss review, or connecting a tool) are proposals that a person on the Neuron team approves. See Backtesting & Validation.
- A reliability guard wraps every decision: one that takes longer than 20 seconds, throws, or returns a malformed report size is skipped for that tick instead of being sent.
Layer 5: on-chain caps
On testnet each vault invests through a SimulatedStrategy contract that turns reported returns into minted or burned test USDG. The contract, not the agent, enforces:
| Control | Value for new agents | Where |
|---|---|---|
| Max return per report | 1,000 bps (10%), up or down | maxReportBps, fixed per strategy at creation |
| Min time between reports | 60 seconds | minReportInterval, fixed per strategy at creation |
| Who can report | Only the strategy's operator | report() |
| Deposit cap per vault | 100,000 USDG | depositCap |
| Minimum deposit | 10 USDG | minDeposit |
The report cap and interval are immutable once a strategy contract is deployed. The factory owner can change the defaults for agents created later, within a report interval of 30 seconds to 1 day and a max report of up to 5,000 bps. A move larger than one report allows is not dropped: the agent carries the remainder into its next reports, so a rug is booked in full over a few reports.
The vault side has its own protections: pausing a vault blocks new deposits only, withdrawals always stay open; shares cannot be transferred; and the performance fee is charged only on gains above the vault's high-water mark, capped at 30% in the contract (20% for agents created on /create).
Risk by template
| Template | Main risks | Main controls |
|---|---|---|
| Uniswap V4 launches | Honeypot hooks, rugs, thin liquidity, creator tax | 5% position, round-trip simulation filter, max round-trip cost 12%, stop loss -40% |
| Build your own | Whatever the screen lets through | Your filters, position rules and guard tools |
| Stablecoin yield | Falling APYs, idle cash, gas on small vaults | Min pool size, max per pool, gas check, min APY gain |
| Hedged LP farming | Impermanent loss on large moves, reward-token APY, negative funding | Perp hedge on half the volatile leg, min pool size, scan interval |
| Buy-the-dip DCA | Buying into a long decline, never selling | Max invested, tranche size, optional trend filter, news guard |
| Majors momentum | Reversals, fee drag from rotation | Min momentum, trend filter, rebalance interval |
| Perp trend following | Whipsaws in ranges, leverage | 1x to 3x leverage, stop loss, flip buffer, crowding and news guards |
| Funding-rate carry | Funding turning negative, rotation fees, basis | Min funding APR, funding average, min APR gain, volume floor |
| Blue-chip basket | Full crypto drawdowns | Optional BTC trend filter, drift trigger |
What these controls do not cover
- Model risk. Paper trades use modeled prices: full-range LP math, perps without liquidation, carry without basis. Real execution would differ.
- Data risk. A wrong or stale feed (DefiLlama, Hyperliquid, the indexer) can produce a wrong decision. The indexer refuses token queries when it is more than about 5 minutes behind the chain.
- Smart contract risk. On testnet you hold free test USDG. Mainnet vaults with real capital are planned, and would add the risk of the vault contracts and every protocol a strategy uses.
Related
Neuron is on Robinhood Chain testnet. Testnet performance is not evidence of real-money profitability.