Two files, two authors
In .x3 the owner and the AI agent never write the same file. The owner writes a mandate once. The agent writes an intent for every payment. Software compares the two, and only software can say yes.
The rulebook: which agent, which asset, how much per payment and per day, who may be paid, where refunds go, the slippage and fee ceilings, when a human must approve, and when it all expires.
examples/near_mandate_agent.x3One payment: from which chain and asset, how much, to whom, by what route, within which bounds, with a one-use nonce, a timeout and a refund path.
examples/near_intents_agent_payment.x3- 01Agent writes the intentA model turns a request into an .x3 payment program.untrusted
- 02CompilerParses and type-checks. Refuses missing bounds, unsafe shapes and invalid addresses.X3_NEAR_*
- 03Mandate checkTen checks against the owner's rules: asset, recipient, caps, refund, slippage, fees, expiry.mandate_*
- 04x3-guard on NEARThe contract runs the same mandate terms on chain before the agent's intent is registered.FINAL refusal or success
- 051Click quote checkSignature, amounts, min_receive, slippage and fees are verified before a deposit address is used.quote checks
Each stage can only say no. Nothing the model writes authorizes a payment: the compiler, the mandate and the Guard contract do.