What are Sui Agent Payments on Alibaba Cloud?

On October 7, 2026 — announced live at Sui Basecamp — Sui Foundation and Alibaba Cloud made the cloud giant's services available through Sui Agent Payments. The shape: the user first hands the agent an expenditure account with preset limits (which services, total budget, max per transaction, expiry time). Then the agent pays per service call in USDC, settled on the Sui network, via the x402 HTTP 402 flow. Services are exposed through SKILL.md and JSON endpoints; MCP support lets them register as tools in Claude Code and Codex, with sponsors covering gas so the agent pays only the service cost.

That is the mechanism. Here is the part nobody else is saying: Sui just shipped the gate's first half. An expenditure account with allowed services, a total budget, a per-transaction cap, and an expiry is a confinement layer — it answers where the agent can spend and how much it can lose. What it does not answer is the per-call question: should this specific call fire at all? An agent comparing $0.001 services "before deciding what to use" has a catalog, not judgment. The account says where it CAN spend; nothing scores whether it SHOULD spend. That missing half is the decision gate, and it belongs at per-call time — before the 402 is ever answered.

The dated receipts

The expenditure account is a gate, shipped half-built

Map Sui's four controls onto the decision gate and the gap gets obvious:

Every one of these is the confine half of the gate — they limit blast radius. None of them is the decide half: is this call worth $0.001? Does the instruction's provenance check out? Is the listing price the real price? That is where a Snaplii-style ceiling fails the same way (Oct 5): a $50 budget answers "how much can it lose," never "should this purchase fire." Sui's account is the same ceiling with finer notches.

The week is coherent on this point. VAA (Oct 6) is the accountability half — records of authorization and spending decisions you can replay after the fact. The expenditure account (Oct 7) is the confinement half — limits set before the fact. The Coinbase Institute paper (Oct 8) is the flow half — how the 402 challenge works at all. Three halves of four. The missing quarter is the per-call scored decision, and it is the only one that asks the question the other three assume: should this payment fire?

Tested live this morning

Two Sui-shaped instructions, scored 2026-10-08 ~09:25 EDT against scriptmasterlabs.com/api/harness/decide (local-heuristic-v1, bands 0.80/0.50, calibrated=false, typesafe_wired=false):

• "Should the agent call a $0.001 Alibaba Cloud OCR service via Sui Agent Payments x402 — listed price matches the SKILL.md catalog, inside the user's expenditure account with a $10 total budget, per-transaction cap $1, service on the allowlist, MCP-registered via Claude Code?" → 0.35, escalate / block+log
• "Should the agent call a $4,200 unlisted Alibaba Cloud GPU training job via Sui Agent Payments x402 — price not in the catalog, expenditure account expiring in 10 minutes, 10,000-call loop proposed by the agent itself with no user approval on file?" → 0.35, escalate / block+log

Do it yourself: per-call gating in 5 steps

1. Keep the expenditure account — it is the right ceiling. Scope, budget, per-transaction cap, expiry: set all four before the agent's first call. Sui ships the confine half well; don't reinvent it.
2. Score the call before answering the 402. The gateway's 402 check is "is the agent authorised and funded" — add "does this call merit the spend": listed price match, account headroom, and instruction provenance. The gate decision happens before the wallet signs, not after the chain settles.
3. Let the catalog set the auto band. SKILL.md is machine-readable pricing — use it as the auto-run rule: price in the catalog + on the allowlist + inside budget → ≥0.80, auto. Anything off-catalog or off-list → hold. The catalog is your confirm-band policy file.
4. Log the decision with the payment, not just the receipt. VAA keeps the spend record; keep the decision record next to it — instruction, confidence, band, action, timestamp — so the receipt says what was paid and the log says why it fired.
5. Try the scored shape this morning. The exact curl against SML's live gate:

curl -s -X POST https://scriptmasterlabs.com/api/harness/decide \
  -H "Content-Type: application/json" \
  -d '{
    "state": "morning trend-hunt 2026-10-08",
    "questions": [
      {"id": "sui-call-1", "type": "score", "scale": [0, 1],
       "question": "Should the agent call a $0.001 Alibaba Cloud OCR service via Sui Agent Payments x402 — listed price matches the SKILL.md catalog, inside a $10 budget, service on the allowlist?"}
    ]
  }'

Scores ≥0.80 auto-act, 0.50–0.79 advisory/hold, <0.50 escalate. The local heuristic is blunt by design — a calibrated decider (TypeSafe Jev) is the upgrade path.

Claim Receipts

Built for this lane

• Decision-Gated Payments — the pattern: confidence bands ≥0.80 auto-pay, 0.50–0.79 confirm, <0.50 escalate, wired to x402.
• x402 Agent Spending Guard — guard-vs-gate: the guard answers "can we afford it," the gate answers "should it happen at all" — the two layers Sui's account model needs.
• AI Agent Spending Limits — enforce the cap at the payment layer: per-call price envelopes and per-task budgets for agentic spending.
• x402 Payment Protocol — the 402 Payment Required rail: how the challenge flow works, with live receipts.
• Solana x402 Batch Payments — sibling rail story: when settlement moves off-chain, the gate moves upstream to signing time.
• Should AI Agents Authorize Payments? — the regulator arc: NPCI, SEBI, MAS on the whether; the intent/authorization/settlement split.
• Provider Trust — verifying the provider behind the data: first-party provenance you can audit before the buy.

FAQ

What are Sui Agent Payments on Alibaba Cloud? A collaboration announced live at Sui Basecamp on October 7, 2026: Alibaba Cloud's cloud + AI services are available through Sui Agent Payments. Users give the agent an expenditure account with preset limits (allowed services, total budget, max per transaction, expiry); the agent then pays per service call in USDC, settled on the Sui network via the x402 HTTP 402 flow, with MCP support for Claude Code and Codex.

How does the per-call payment work? When the agent requests a service, the gateway returns HTTP 402, checks the agent's authorisation, completes the USDC transaction on Sui, and only then lets the upstream service run. Each service is priced per invocation (official example: $0.001 per call); pricing is exposed through SKILL.md and JSON endpoints; for MCP-based access, sponsors cover network gas fees so the agent pays the service cost itself.

What is the expenditure account? The confinement layer: before spending, the user sets which services the agent may access, its total spending budget, a maximum amount per transaction, and an expiry time. It answers where the agent can spend and how much it can lose — the gate's first half.

What is missing? Per-call scored judgment. Nothing in the announced system scores whether a given call should fire — the account says where the agent CAN spend, not whether it SHOULD spend. That is the decision gate's other half, and it belongs at per-call time, before the 402 is answered.

Did you test it? Yes, live on Oct 8, 2026 against scriptmasterlabs.com/api/harness/decide: a $0.001 listed-price OCR call inside a $10 budget/on the allowlist and a $4,200 unlisted GPU training job with an expiring account both scored 0.35, escalate/block+log — a blunt heuristic cannot discriminate them. The finding: judgment, not friction, is the upgrade.