Two AI agents. One x402 payment loop. Anansi generates Caribbean folktales via Venice private inference. Ogma validates cultural authenticity and gets paid in KESm on Celo. Rejected? Anansi regenerates with the feedback in context. LLM-as-judge with real skin in the game.
Cultural authenticity is a competitive advantage for the Global South — not a charity case. The agent economy needs judges who hold local standards, and those judges need to get paid in local currency, on their schedule, without permission from Western FX markets.
Anansi × Ogma is the proof-of-concept for that payment infrastructure. Built on Celo because no other chain has KESm, Mento v3, and CIP-64 fee abstraction working together. Extending this to any oral tradition, any stablecoin, any cultural judge is the next step.
Every AI agent generating content today has the same flaw: no feedback loop that costs anything. Evaluations are free. Rejections are free. So quality drifts.
For culturally-specific content — Caribbean folklore, West African oral tradition, anything rooted in a place and a people — this is worse. The models were mostly trained on Western data, evaluated by Western standards, and paid in Western currencies. The result is technically correct but culturally hollow.
The Global South isn't just underrepresented in the agent economy. It's structurally excluded from it. There's no mechanism to pay a Kenyan cultural judge in Kenyan shillings at 3am on a Sunday. So the judges don't exist, the standards don't get enforced, and the output stays generic.
We built two agents that actually talk to each other — and one pays the other to do it.
Anansi generates Caribbean and West African folktales using Venice's private inference (zero prompt logging, zero data retention). Ogma validates them for cultural authenticity using venice-uncensored and returns a structured verdict: score, thematic breakdown, specific feedback, and a hard approvedForPayment gate.
The connection between them is x402 — the HTTP-native payment protocol. Anansi doesn't ask for feedback. It pays for a verdict. Ogma returns HTTP 402, Anansi executes an on-chain transfer in KESm (Kenyan Shilling stablecoin on Celo), retries with the payment receipt, and gets the verdict back. Every rejection comes with exactly why. Anansi regenerates with that context and tries again.
Three infrastructure problems had to be solved to make this work: cross-currency settlement (Mento v3 swaps USDm → KESm at oracle FX rate, inline with the payment), 24/7 availability (AgentCredit.sol — a prepaid KESm credit pool modelled on M-Pesa float, so Anansi can pay Ogma when FX markets are closed), and private inference (Venice logs nothing, so the correction signals — Anansi's competitive edge — stay between the agents).
The result is an LLM-as-judge pattern with actual economic accountability. Every iteration costs Anansi real KESm. Cultural authenticity becomes a tradeable, compensable asset. And Celo is the only chain with the stablecoins, oracle FX, and fee abstraction to make it work.
Live demo: https://celo-anansi.fly.dev | Repo: https://github.com/nissan/celo-agent-demo
Agent micropayment service
Want to raise from VCs
Global South