AI-assisted swaps: where automation stops and settlement begins

An assistant can help prepare a swap, but settlement starts somewhere else: with the transaction you review and sign in a wallet. That distinction matters because a draft, a quote, and a confirmed ledger row answer different questions.

Five stages, five kinds of evidence

Think of an AI-assisted swap as a handoff between stages:

Stage	What it can establish
Prompt or model answer	A suggested explanation, comparison, or checklist
Review plan	The action, route, and inputs you intend to inspect
App quote	Current output, minimum received, fees, route, and slippage shown by the app
Wallet approval	The exact transaction you chose to authorize
Confirmed readback	A digest and service record that the ledger has verified

The first two stages are preparation. They can expose a missing amount or an ambiguous asset type, but they cannot make a transaction real. The quote is time-sensitive information. Wallet approval is an explicit authorization. Only a confirmed and verified record can support an activity credit.

Use a read-only plan first

The participant agent kit exposes six bounded tools. suibrokers_get_status reads health and public runtime configuration. suibrokers_get_season reads one published stage. suibrokers_get_swap_history reads a wallet-scoped history when you provide a canonical address. suibrokers_get_desk and suibrokers_get_partner read their public records. suibrokers_prepare_action returns a review checklist and an app route; it does not quote, build, sign, send, activate a referral, or mint an NFT.

That makes a useful first request quite specific:

Prepare a review-only swap plan for this wallet and these two Sui coin types. List the values I must check in the app. Do not quote, build, sign, send, or claim Points.

If a tool returns unavailable, keep that state. Do not turn an unavailable history into an empty history, and do not fill a missing quote with a remembered price. The Swap skill describes the same boundary and the fields to inspect.

The quote is still the source for the decision

In the app, check the connected network, input and output asset identity, amount, route, quoted output, minimum received, slippage tolerance, gas, and any partner commission. A commission and network gas are separate fields. A model can explain those labels or calculate a difference between two quotes that you pasted, but it cannot make an old quote current.

For example, give the assistant two quotes from the same pair and ask it to compare only the fields present. If Quote A has a larger output but Quote B has a larger minimum received, the answer should leave that tension visible. Ask which missing expiry or route detail prevents a fair choice. Then read the live app and wallet prompt yourself.

Signing changes the evidence

The wallet prompt is the boundary where you decide whether to authorize. Check the recipient, assets, amounts, network, gas payment, and any contract interaction. A model summary is not a substitute for the prompt. If you reject the transaction, there is no settled swap to credit. If it confirms, wait for the service readback rather than treating a success message or a copied digest as proof.

The current Swap rule credits one Point per $4 of verified volume. A verified $100 example corresponds to 25 Points; it does not predict a return or guarantee an NFT. The Points and Broker NFTs guide explains why a threshold, qualification state, capacity, and final allocation remain separate records.

Where automation stops

An assistant is good at turning a supplied quote into plain language, finding a missing field, and keeping a review journal. It can also call a read-only function when a host and tool server are configured to allow it. OpenAI’s function-calling documentation describes the model requesting a function while the application executes it and returns the result. That is a useful model for understanding the boundary: tool access belongs to the application, while wallet authority belongs to the person who approves the transaction.

Keep the raw quote, wallet prompt, digest, and confirmed history together. That small habit makes it possible to distinguish an assistant’s suggestion from a settled event when you review the activity later.