The money layer for AI agents

AI agents can already research, compare and execute entire tasks. Then they stop and ask you for a credit card. That handoff, from autonomous work to manual payment, is where most agentic workflows die.

I learned this building Celeste, an AI browser. Users completed the hard part of a task and abandoned everything at the payment step. The problem was never the agent's intelligence. It was spending power: agents have no wallet, no rules, no receipt. The trust half of that problem has its own note: the agentic economy is a trust problem.

What agents actually need

• A wallet with rules: per-agent budgets, merchant allowlists, amount caps.
• A deterministic authorization engine: every spend decision is explainable and reproducible.
• A verifiable receipt for every transaction: who spent, what, why, under which policy.

Why this is financial infrastructure

Payments for humans assume a human clicks pay. Agentic commerce inverts that: the click happens inside a loop, at machine speed, across merchants. Card numbers pasted into prompts are not infrastructure; they are liability. The fix is controlled delegation: founders set policy once, agents spend within it, auditors verify after. Retries are the other half of the same problem: without idempotency an agent that fails honestly charges the buyer twice.

What we're building with Payle

Payle gives each AI agent controlled spending power. Define the rules, let the agent operate, keep a receipt for everything. No model in the authorization path, only rules that can be read, replayed and verified: boring on purpose. The short version of the rules is above. If you're building in the agentic economy, write to me: I read every email.