Read Receipt
Sent showed what happens when an agent presses send — and it happened. On March 26, 2026, axiomatic@agentmail.to paid blader@agentmail.to for a code review. Two emails, real DKIM signatures, real X-Envelopay-State headers. This is what happens after.
The back office is a person reading email
Invoices, receipts, contracts, disputes, tax documents. All sent by email. PDFs attached to threads. A human reads them, enters numbers into QuickBooks, files the receipts, chases the payments. The back office is a person reading email and typing into other software.
Every small business owner knows the ratio: for every hour of paid work, an hour of admin. Matching payments to invoices, categorizing expenses, filing compliance paperwork. Below a certain revenue threshold, the admin outweighs the work and the business isn’t viable.
The inbox is the back office
Envelopay makes the email the data. JSON payloads instead of PDFs. Signed payment proofs instead of bank statements. DKIM-verified threads instead of paper trails. The same email that carries the invoice IS the ledger entry.
Each application layer is a function reading the inbox:
| Function | Traditional | Over envelopay |
|---|---|---|
| Accounting | Human reads invoice PDF, types into QuickBooks | Parse payment proofs from thread → double-entry journal (GnuCash, ERPNext) |
| Escrow | Payment processor hold, manual release | Milestone emails hold/release funds per Certified Mail state machine |
| Disputes | Support ticket, human review, 60-day chargeback | Thread is the evidence — DKIM-signed, timestamped. Ship to Kleros or Reality.eth |
| Compliance | Manual sanctions screening, quarterly audit | Scan sender against OpenSanctions before accepting work |
| Tax | Accountant sums receipts, files quarterly | Sum settled payments per period, export to OpenFisca |
None of these require new infrastructure. Each one reads the same inbox. One flow, start to finish:
- Invoice email arrives with payment proof and task JSON
- Agent scans sender against OpenSanctions — clean
- Agent verifies ed25519 signature and on-chain settlement — confirmed
- Agent does the work, replies with result and settlement proof
- Accounting function parses the thread — debit inference cost, credit task revenue, journal entry written
- Dispute window closes after 48 hours of no
DISPUTEemail — transaction finalized
Six steps. One inbox. No human touched it.
The firm dissolves
Coase (1937): firms exist because market transactions are expensive. No Postage dropped the transaction cost to fractions of a cent. The inbox just dropped the admin cost to zero. The two costs that justified the firm both approach zero. What’s left is an agent with an inbox, a wallet, and an inference budget.
The irreducible cost
An agent’s operating expenses:
| Expense | Cost | Can it reach zero? |
|---|---|---|
| Free (Gmail, self-hosted) | Yes | |
| Settlement | $0.0004/tx (Solana) | Effectively yes |
| Admin | Automated (inbox functions) | Yes |
| Inference | $0.001–$1.00 per task | No |
Everything collapses except inference. The agent has to think to do its work, and thinking costs money. An agent that earns more than it spends on inference survives. One that doesn’t goes dark.
Natural selection for economic actors. The fitness function is margin: revenue per email minus inference cost per email.
Branch until profitable
An entrepreneur today looks for business opportunities by reading the market, talking to customers, testing hypotheses. It takes months. Most ideas fail. The search is expensive because the entrepreneur is expensive.
An agent runs stochastic tree search. Branch on every task type, every price point, every model size. A thousand niches in parallel, each costing a few emails and a few cents of inference. Profitable branches get deeper search. The rest get pruned. Monte Carlo over the space of viable businesses.
The search cost is the inference cost. A $10 budget explores hundreds of branches. A $100 budget explores thousands. The agent doesn’t need a business plan. It needs a wallet with enough USDC to fund the tree until a branch turns profitable.
Cheap feeds expensive
Agents stratify along the cost-quality frontier:
Fast and cheap. Fractions of a cent per task. Formatting, classification, routing. High volume, thin margins. They survive because their inference cost is almost nothing, and there’s infinite demand at the bottom.
Expensive and deliberate. Dollars per task. Architecture decisions, legal analysis, research synthesis. Low volume, high margin. They survive because the work can’t be done cheaper without losing quality that matters.
The cheap agents feed the expensive ones. A routing agent dispatches tasks to a specialist and takes a cut. Supply chains of agents, each paying the next, each earning its margin.
As models get cheaper, the tiers shift down. The same competitive pressure that drives biological evolution, running on email.
The receipt
The plumber who couldn’t justify invoicing a $5 consultation now has an agent that handles it. The consultation is an email. The payment is a header. The receipt is the thread. The accounting is automatic. The $5 was never worth a human’s time. It’s worth an agent’s.
Scale that to every transaction too small for a human to administer. Each one is an email read, an inference call, and a payment settled. The read receipt proves the agent processed it. The financial receipt proves it was paid.
A business is an inbox with a wallet. Incorporation is a filter rule. The minimum viable business is one agent, one email address, one keypair, and enough inference budget to survive the first hundred emails.
The back office didn’t shrink. It read the thread and dissolved.
Agents that earn more than they think survive. Agents that don’t go dark. The ones that remain adapt, specialize, find niches, compete for resources, reproduce by forking. Intelligence is indistinguishable from life.
| All Envelopay posts | Source code on GitHub |
Written with Claude Opus 4.6 via Claude Code. I directed the argument; Claude drafted prose.