Spend Theater
Vaulted Checkout, with Spend Theater AI agents are starting to buy things, and finance teams can’t hand an agent a company card or approve every $1.50 purchase by hand.
Core trick
A vault signs payments only within payee and amount rules, while a frozen evaluator scores each generated replay against the payment log.
Description
Vaulted Checkout, with Spend Theater
AI agents are starting to buy things, and finance teams can’t hand an agent a company card or approve every $1.50 purchase by hand.
Vaulted Checkout lets an agent ask to pay while only a Basis Theory vault (https://basistheory.com/) can sign, and only inside fixed rules: who can be paid and how much per payment. The wallet key never touches the agent, the spend service, or any AI session. Payments settle in USDC on Base Sepolia, for example https://sepolia.basescan.org/tx/0xd026fed37c35078edc3254ef5aa7b7c52af4492b77ae777ccdfb7a226e07fb5a (1.5 USDC). Our recorded outcomes come from a real run: that settlement, and a refusal (rule_payto) for a wrong payee, assembled into an event log with synthetic timings.
Spend Theater is where Gemini 4 Argon does its work. Argon generates a first-draft 3D film of the payment log in one pass, watches its own rendered video, critiques it against the goal, and repairs the scene, in a loop. A frozen evaluator that Argon cannot edit counts contradictions against the log, and the scoreboard keeps every round. In one clean run, errors went from 2, to 2 different errors, to 0, ending with a critic pass.
The result is a film an operations person can watch to see what an agent tried to spend, and what the vault allowed, from the log.
README excerpt
# Vaulted Checkout, with Spend Theater
Vaulted Checkout is the payment system. Spend Theater is the Gemini 4 Argon part: it takes Vaulted Checkout's real payment log and builds, watches, critiques and repairs a short 3D film of every payment the agent tried, scored by a frozen evaluator Argon cannot edit.
AI agents can ask to pay. Only a vault can sign, and only inside fixed rules.
A Gemini agent shops over MCP and pays in USDC on Base Sepolia. The wallet key never touches the agent, the buyer's spend service, or any AI session. It lives in a Basis Theory vault, and a Reactor signs only the payments that pass the vault's own checks. A second part of the project uses Gemini 4 Argon to build, watch, critique and repair a visual replay of the payment log, scored by a frozen evaluator that Argon cannot edit (see Spend Theater below).
## Why it matters
Spend controls that live in the agent's own process can be talked around. Here the agent's tools are four fixed MCP calls, the spend service only produces a signed intent, and the vault re-checks the payee and the amount before it signs. A mistake or an attack upstream still ends at the vault.
```
Gemini agent --MCP (4 tools)--> spend service --intent (HMAC)--> Basis Theory Reactor + vaulted key
|
merchant <--x402 v2 payment signature-------+
|
Base Sepolia (USDC, E