When the agent touches money,who authorized it?
Every finance team running AI agents is making economic decisions without a paper trail. Logs say something happened. KIFF proves what was authorized — before the money moved — in a way nobody can alter afterward.
first protected control: duplicate-payment guard
$10,000 invoice.
One payment.Nine blocked retries.
An AP agent pays an invoice. A flaky connection drops the success response. The transport retries ten times. Each retry is a legitimate call — same invoice, same amount. Without KIFF, that's $100,000 across ten debits. Only a state-aware gate can stop it.
Every individual $10,000 call was legitimate. Only a state-aware gate stopped the repeat. The state machine is not a dedup table in the application — it is an independent authority that knows Invoice inv-001 was PENDING when the first call arrived, and PAID when the retry arrived. That fact is recorded, signed, and tamper-evident. That is not a log. That is Exhibit A.
Six phases.One protocol.
Same for humans, agents, services.
Every action — human or AI — follows the same clearance path. No agent gets special treatment. No action escapes the record.
Six actions that move money.All need Exhibit A.
Start with whichever action you are most afraid to let an agent execute without a guard. Each becomes one Protected Control — one production action KIFF decides before execution and proves afterward.
Logs say what happened.
Receipts prove what was authorized.
The protocol is a Go package.
Embed it in your backend.
The framework is public, MIT-licensed, and framework-agnostic. Wire it into an existing Go backend in an afternoon. Or connect your agent runtime to KIFF Cloud without touching your application code.
What is the first costly action
you don't want an agent to execute without a guard?
Start with the sandbox to prove KIFF can observe your runtime and draft a Protected Control. When you're ready to put it in production, the Launch Pilot activates it with founder-led setup.