.png&w=828&q=75)
Most "autonomous" purchasing agents are safe because their prompt tells them to be. Limen removes that assumption. The per-deal ceiling is written into an escrow contract by a human, and the agent has no path to change it. Instructed to overspend by fifty dollars, the transaction reverts with ExceedsPerDealCap. Instructed to raise its own cap first, that reverts too, because the policy record keys off the message sender and an agent can only ever write its own. The work moves through three desks that cannot do each other's jobs. Sales describes what is needed and reviews what the agent found. The Head of Operations approves the amount, publishes the ceiling on-chain and funds the escrow. Finance releases payment and cannot approve. Roles are read from a signed token, never from the browser, and every refusal is a real HTTP failure rather than a hidden button. A language model explains the deal in plain English on the approval screen, but it is held outside the money: a rewrite that drops, alters or invents a figure is rejected against the deterministic text before it reaches a person. That boundary is enforced in code, not in a prompt. The demo runs a full purchase: screening fifteen suppliers, three negotiation rounds, approval, escrow, a two-party delivery attestation, and settlement with a signed invoice and an exportable audit trail. Payment top-ups run through Razorpay in test mode. Built by Sujal Negi, a student at IIITDM Kurnool.
1 Dec 2026