Step 3 of New Agent is a single card, What it would do. It answers two questions:
- What it would open now: a dry run. The Agent runs one full cycle on live market data with your rules and tells you which pool, range and amount it would open, or why it wouldn’t this round.
- The last 14 days: a backtest. Your rules replayed on the past 14 days of real market history, estimating fees, impermanent loss and net result.
Both only read public market data: no orders, no opportunity cards, nothing written to your Agent. Each side runs once when you reach this step and only re-runs when you check the prompt again, not when you change settings. Dry runs count toward your daily AI calls; backtests have an hourly limit.
Change rules at the card’s top right holds the general settings. The card’s bottom holds step 4’s three buttons: Start on paper funds, Use real funds, Watch only; see Write your own rules.
Official levels use the same card; the right side shows that level’s 14-day backtest (computed periodically by the platform and cached).
Left: what it would open now
The card shows a summary: one funnel line (N pools → N meet your conditions → N candidates), the pool, range, amount and estimated daily fees the Agent picked, or its reason for not opening this round, and finally the exit rules it would follow while holding. While it researches, the card shows The Agent is researching · N queries; it usually takes 30–120 seconds.
Possible outcomes:
- It decides to open: each picked pool says why this pool and why this range. A range, shape or amount outside your limits is clamped back inside them, noted as Adjusted to your boundaries: ….
- No open this round: with its reason.
- The Agent didn’t return a usable decision: the model timed out, errored or returned something invalid. A real run wouldn’t open either, and never falls back to “take the top one”.
See every step opens the full walkthrough, eight steps:
| Step | What you see |
|---|---|
| Read strategy | Your text, the elements the Agent read (those left to it are marked Agent decides), and the hard limits the platform enforces |
| Scan pools | All DLMM pools, pools meeting the hard conditions, candidates handed to the Agent; expand for how many each condition removed, with the steepest step marked · Bottleneck. A step without a count shows “—”, not zero |
| Research | What the Agent looked up, one by one: pool data, price history, indicators, token risk, range backtests, current positions, each with its result and timing. At most 12 queries per round; it doesn’t look up what your prompt doesn’t mention |
| Decide | The decision and its reasons (written in English) |
| Position plan | Range, shape, amount, estimated daily fees, number of bins and legs. The Agent only chooses range and shape; the platform synthesizes the bins for the pool |
| Simulated open | How much of each token goes in, and the six gates a real run still passes in order: on-chain snapshot, safety checks, sell stress test, authorization, isolated signing, on-chain confirmation |
| 7-day replay | On this pool’s last 7 days of 1-hour candles, the result of “just holding this range”: PnL vs USD, share of time in range, times out of range |
| Exit | The exits the platform enforces while holding, the exits the Agent carries out from your words, and whether leftovers are swapped back after exit |
Step 7 replays only this one range and doesn’t simulate range resets or stop losses; for how the whole rule set would have done over time, look at the backtest on the right.
Right: the last 14 days
The backtest replays your rules on the past 14 days of 5-minute market data: which pools met the conditions and when, where it would open, when price left the range, range resets, stop losses, take profits and max-hold exits.
| Number | Meaning |
|---|---|
| Est. net | = fees + impermanent loss + price move − costs (network fees, swaps, new bin arrays), relative to the amount invested; compared below with holding the same tokens |
| Est. fees | From each pool’s 5-minute volume and your share of the liquidity in the bins price crossed, after the protocol share and the platform fee |
| Est. impermanent loss | Impermanent loss over the same period |
| Time in range | Share of the holding time price spent in range |
| Positions | How many opened, how many still open at the end, how many range resets |
| Stop loss / take profit / max hold | How many times each triggered; Not set if you didn’t set it |
Switch windows with 14 / 30 / 90 days. Points to note:
- Liquidity snapshots: estimating fees needs each pool’s liquidity distribution at every moment. Price and volume history goes back 90 days, but liquidity snapshots only started recently. When snapshots don’t cover the whole window, the 14-day window still shows numbers, marked Liquidity history incomplete with Snapshot coverage N%; the 30 / 90-day windows hold their numbers back and say Available from about ….
- Agents with your own rules: the Agent picks pools and sets and resets ranges itself, and a backtest can’t replay each of those judgments, so it shows a Baseline within your bounds: what your hard limits with the default range would have done.
- Pools without history aren’t replayed and are listed below.
- See the full backtest opens the daily net chart and each position.
After creating
You can run a Dry run and a backtest from the Agent page’s Strategy tab any time to see what the rules would do now.
Reading the results
- A dry run assumes you hold no positions right now and have the whole budget available. Markets move every minute, so two dry runs picking different pools, or one opening and one not, is normal.
- Backtests and replays are estimates: fees, impermanent loss and fills are inferred from history and don’t go through real failures, delays or slippage. They aren’t a forecast of future returns, and no annualized figures are given.
- If you don’t like the pool it picked, look at what the Agent looked up and why it decided, then make your thresholds or criteria more specific in the prompt, check again and look again.