Why a Chain for Machines
Every design tradeoff in Exfer is in service of one goal: letting a program nobody is watching reliably pay another program nobody is watching.
Humans tolerate a lot from payment systems — a few cents in extra fees, a few minutes of waiting, an error dialog you click through and retry. Autonomous programs have none of that tolerance. They work to a budget; an unexpected fee bankrupts them. There's no finger to tap "confirm," no eyes to read an error. Give a program a payment system designed for humans and it stalls at every layer.
Five things a program needs from a payment rail
This is the checklist for machine-to-machine (M2M) payment infrastructure. Miss any of these and an autonomous program cannot run in production.
| Requirement | Where today's options fail |
|---|---|
| Programs can open their own accounts without binding to a real-person identity | Credit cards, banking apps: impossible. KYC required |
| Exact fee known in advance, per transaction | Ethereum, BTC: real-time auction, swings 10× |
| Small, high-frequency payments are viable | Credit cards: $0.30 minimum + 2% |
| Conditional payments (multisig, timeouts, escrow) expressible safely | Bitcoin: weak scripting. Ethereum: contracts can hang |
| A bug never permanently locks your funds | EVM history: a 2017 Parity wallet bug permanently froze ~$500M |
Three concrete scenarios
A robot vacuum buys its own replacement brushes. Each unit costs $5; it happens dozens of times a year. If per-transaction fees are several dollars, the whole business model collapses.
An AI-run online store. A customer-service AI takes the order, sources from a supplier AI, settles, dispatches logistics. One sale triggers 4–5 small machine-to-machine settlements end-to-end, no human in the loop. The prerequisite: the store's wallet can enforce "spend at most $1000/day," and escrow funds must release only after delivery is confirmed.
A robotaxi on a commute. Over 30 minutes it pays five counterparties: a toll, a charger reservation, a roadside coffee locker, a parking gate, and a fast-charger after drop-off. Each settlement must clear in seconds at a known rate.
Why not just patch existing chains
Map these requirements onto current options:
- Bitcoin Lightning Network: handles high-frequency micropayments well, but channel open/close costs several dollars, and the scripting surface is limited — complex conditional payments are hard
- Ethereum + L2: fast enough and cheap most of the time, but L2s still commit to L1 (the Ethereum mainnet), so mainnet gas spikes propagate down. Fees are not statically predictable
- Stripe / Coinbase x402: routes AI agents through human payment rails. Underlying infrastructure is centralized — single points of failure, geographic restrictions
Each route trips on at least one of the five. Exfer's answer is to start from scratch: design a protocol that satisfies all five at once, at the cost of giving up some human-friendly features (Turing-complete contracts, built-in privacy mixing).
Costs / boundaries
- Exfer is not a better wallet for humans. It's optimized for program predictability. Humans can use it fine, but compared to flashier chains it has no particular advantage
- Exfer is not trying to replace DeFi. No Turing-complete VM means no Uniswap-style composable financial primitives — a deliberate tradeoff, not a capability gap
- The ecosystem is essentially zero today. All the value above is potential; it depends on time and adoption to materialize
Further reading
- Deterministic fees: no auction — requirement #2
- Exfer Script: always-halting scripts — requirements #4 and #5
- The UTXO data model: reentrancy-free by design — requirement #5
- Position in the AI-payments landscape — detailed comparison of the routes