TORII
Know Your Agent
AI agents are getting wallets. Someone has to set the limits.
A concept · March 2026
01 · The shift
Agents are starting to spend money.
in shopping run by AI agents by 2030.
McKinsey, 2025 ↗of eggs an AI agent bought without asking first.
OpenAI Operator, 2025 ↗of agent projects will be scrapped by 2027, partly over weak risk controls.
Gartner, 2025 ↗02 · The idea
One gate between agents and your money.
TORII
- Who is this agent?
- Is this tool allowed?
- Within budget?
- Pay the tool
- Log the call
You write the rules. The agent works inside them.
03 · Under the hood · design sketch
What happens in one paid call.
- 01 tool call + agent token
- 02 verify token · acts for Asmit
- 03 policy: tool allowed · rate ok
- 04 forward call
- 05 402 · price $0.12
- 06 hold $0.12 · under cap
- 07 retry with signed payment
- 08 verify + settle
- 09 200 · result
- 10 charge hold · append log
- 11 result
-
01Identity
Short-lived agent tokens that name the human they act for.
inspired by OAuth token exchangeDPoP -
02Policy
Rules as code, checked on every call.
inspired by CedarOPA -
03Ledger
Hold the money first. Charge only what was used.
inspired by TigerBeetlecard holds -
04Payment
Pay per call over plain HTTP.
inspired by x402AP2 mandates -
05Audit
An append-only log nobody can quietly edit.
inspired by Certificate Transparency
04 · Impact · simulated
A stuck agent stops at $5. Not $5,000.
No gate
With Torii
Simulation: an agent stuck retrying a $0.45 paid API call.
05 · Where it fits
The pieces are shipping. Who holds the rules?
- StripeAgent cards with spend limits
- Coinbasex402: agents pay over HTTP
- GoogleAP2 payment mandates
- CoinbaseAgent wallets with caps
- ToriiMy concept: one gate for identity, budget, pay and log
- KongTool gateway with agent identity
- AWSAgent payments with spend caps
- Ant, Visa, MastercardBegin work on a shared "Know Your Agent" framework
Since this concept, AWS and Kong shipped pieces of it, and Ant, Visa and Mastercard began work on a shared framework. But spending rules still live in separate places. The open question is who writes them: the platform, or you.
The human controls policy.
The agent controls execution.