Agentic Payments • 2026
Give your agent a tab, not your wallet.
Role
Founder, Design Engineer
Timeline
2026 (building)
Team
Skills
Agentic Payments
Stablecoin / USDC
Consent & Trust Design
Design Engineering
Overview
Not an agent that pays your bills. The consent surface for agent money.
Tabs is the consent and control surface for agent money, funded by the stablecoins you already hold. The name carries the thesis: put it on my tab, the USDC balance the agent spends from, and keep tabs, the oversight that is the actual product.
Letting an agent buy airtime or pay a bill is the easy part. The hard part, and the part no one has shipped well, is trust: how a person safely grants, scopes, approves, audits, and revokes an agent’s access to money.
This is what I’m building now, the problem Chidi and Kano were both pointing at.
“The payment is table stakes. The product is trust: how you grant, scope, approve, and revoke an agent’s access to real money.”
The convergence
Everyone is racing to put an agent on your money. All of it bank-rails, US-first.
Through 2026 the giants converged on the same idea, an agent that touches your money, and each one stopped at the same line.
Who shipped what
ChatGPT Finance
Connect a bank through Plaid, read-only. Agentic spend is a year or two out, on US bank rails.
Mercury Command
Conversational account ops for US startups; money-movement is deterministic and click-approve. Fiat, B2B.
Paystack Index
An agent-commerce catalog on local rails, airtime, bills, transfers. A fiat balance, a catalog story, not a consent story.
use-agently
Agent-to-tool payments with a dry-run, then pay. A developer CLI, tool payments, not consumer spend.
The open lane
The next billion hold dollars in stablecoins, and want to spend, not just read.
They all left one lane open. The people most ready to let an agent spend for them, fluent in both crypto and AI, keep their money in stablecoins, live in markets these products don’t serve, and want to spend, not watch a read-only dashboard.
The betThe product isn’t the payment, it’s the consent surface. Design that well, fund it with the USDC people already hold, and settle locally to naira. Stablecoin-funded, emerging-market everyday spend, agent-initiated now, the thing ChatGPT is still a year or two from shipping.
“Spend the USDC you already hold. Auto-settle to naira. The control surface is the product.”
What I designed
Four surfaces turn “an agent with your money” into something you control.
The safety model is settled, and the market agrees on it: the AI is read-only and proposes; a separate, deterministic path moves the money once the human authorizes. The design work is making that powerful, scary thing feel calm. Four surfaces do it.
The consent surfaces
The grant
Not one spending limit, but a real vocabulary: per-category caps, a recipient allowlist, velocity limits, time windows, an approval threshold. The screenshot that proves you understand agent permissions.
In-chat approval
Agents have no native modals. Approving “send ₦25,000 to Mum” has to happen inside the conversation, an inline confirm card, without breaking flow. The standout interaction.
The receipt
A signed, shareable trust object that lands back in the chat: USDC debited, naira delivered, rate and fee, all honest. Chidi’s receipt instinct, applied to agent money.
The control panel
The companion surface outside the agent: balance, active grants, live activity, revoke. Where the human stays in command.
Built on the standard
The consent model maps onto an emerging standard, on purpose.
Designing on the emerging standard makes this the consumer surface for the agent-payment stack, not a cool app.
Three mandatesGoogle’s AP2 structures every agent purchase as three signed mandates: Intent, what you authorized; Cart, the specific purchase; Payment, the amount and funding source. Those are literally my three consent surfaces: the grant is Intent, the approval is Cart, the receipt is Payment. The rail underneath is x402, USDC settled on Base.
The prompt-injection answerA malicious site tells the agent to drain the wallet. The mitigation isn’t a warning, it’s the architecture: the agent never holds the keys, and the spend ceiling, allowlist, and approval threshold are what actually constrain it. The grant is the safety story; the kill-switch is the last resort, shown last.
What’s real
One live slice, real money, honest about what’s mocked.
I’m biased toward shipping something real, because real adoption is the one number worth chasing.
The minimum-lovable slice is airtime: a USDC wallet, one scoped grant, an airtime purchase through a real agent, dry-run, approve, receipt, control panel. It ships as a real MCP server so Claude can actually call it, settling USDC to naira over a live emerging-market rail.
The wider catalog, bills, food, transfers, is designed and honestly mocked. Even a few dozen people buying airtime through their own agent is the defensible proof, and the live demo.
What’s next
A family, not an app.
Tabs doesn’t ship alone. It’s one piece of a system, and each piece leads on a different axis so they never read as duplicates.
Kano is the rails.
Stablecoin value across real corridors, human-initiated. Tabs rides the same x402-flavoured settlement, agent-initiated.
The catalog widens.
Airtime first, then bills, food, and transfers, each a new grant category on the same consent surface.
The agent ecosystem.
An MCP server any agent can call. As the next billion meet their first agent, the safe way to let it spend is the product.