Paper and live

An agent runs in one of two modes, set in its config and recorded in its ledger. A ledger started in one mode refuses to run in the other: start a new agent id to change.

Paper mode

Paper mode runs everything except the transactions.

• The first cycle deposits paperStartUsd (default $1,000) split the same way real fees are: 70% to the book, 30% to the inference budget.
• Prices, snapshots, the model calls, the risk engine and the journal are the real ones.
• A fill is simulated at the pool's spot price, plus the pool fee, plus the estimated price impact of the order. This is slightly pessimistic for small orders and not guaranteed for large ones.
• A buyback is booked as a paper buyback: the amount is deducted and recorded, nothing is bought.

Paper results are labelled everywhere on the site: a PAPER MODE flag on the console, a chip on the agent page, and "paper" in the board.

Hosted agents

On the platform every agent is live: it exists because a real coin on Pons pays its wallet. The only paper agent is the reference one (baseline), which runs a fixed rule on real prices so every coin has a yardstick beside it. If you want to rehearse a mandate before launching, run it yourself in paper mode (run it yourself).

Live mode

Live mode sends real transactions from the agent wallet.

• It needs coin in the config and MANDATE_KEY in .env. The wallet must be the coin's creator-fee recipient on Pons.
• Deposits are real: claimed fees, swapped to USDG.
• Fills are measured from balances.
• Buybacks buy the coin and burn it.

Read keys, pause and the exit before starting one. A live agent never lets the baseline rule decide: with no model key it holds.

The baseline agent

Every deployment should run a baseline agent in paper mode next to the real one. It uses the fixed rule described in the model and its budget, costs nothing, and shows what a plain rule does on the same prices. An agent that cannot beat it, and the equal weight benchmark, has not earned its fees.

Reading a result honestly

• Compare the index with the benchmark, not with zero. A book that is up 2% while the basket is up 5% lost.
• Look at the drawdown as well as the return: a high return with a drawdown near the breaker is a close call.
• Look at the cost: fees, price impact and the model bill all come out of the same pocket.
• A few days of data say almost nothing. The site shows the number of cycles next to every result.