A protocol for a world where everyone has a pro on their side.

For most of modern history the people who got the best deal were the people who could spare a Tuesday afternoon to argue about it. The doctor who knew which line on the bill was wrong. The realtor who knew the closing-cost line items by heart. The corporate buyer with twenty years of vendor relationships in their phone. Everybody else paid full freight.

It was an asymmetry of friction more than information. The information was findable. Sample contracts existed. CPT codes were public. Hotel revenue managers had policy doors. The reason most people lost was that negotiating well took time, expertise, persistence, and a thick skin — and very few transactions are worth a Tuesday afternoon when you have a job and kids.

AI changes this. Not in the way the marketing decks imply, with chatbots that sympathize. In the boring way: friction drops to zero. The expertise becomes infinite. The persistence becomes automatic. The thick skin is the easiest part of all.

What changes when friction goes to zero

When friction is zero, every transaction is up for negotiation. Not just the ones big enough to justify a Tuesday afternoon. The cable bill. The insurance renewal. The hospital co-pay. The hotel rate. The weekend appliance install. Every one of them, in the background, while the customer sleeps.

And then the asymmetry flips. The merchant — the hospital billing department, the dealer's internet sales manager, the hotel's revenue desk — used to be the one with all the expertise. Now they're outnumbered. One billing manager faces ten thousand customers each running a buyer agent that has read every claim adjudication appeal in the public corpus.

The merchant needs an agent too. Not to deceive the customer; to apply policy at scale. To say yes within the bands the CFO authorized, no outside them, and to log every decision so the next quarter's policy can be calibrated against the data.

The substrate that nobody is building

Two agents on opposite sides of a transaction need to do five things together: identify each other, exchange offers, agree on what fair looks like, recall each side's history, and settle the deal once they've agreed.

Five primitives. Identity. Messaging. Oracle. Reputation. Settlement.

Each of those was a separate, decade-long product market in the pre-agent world. Stripe is a billion-dollar settlement company. LinkedIn is a hundred-billion-dollar identity company. Bloomberg is a thirty-billion-dollar oracle. None of them were designed with agent-to-agent commerce as the use case. None of them are easy for a small team to compose into a working buyer agent on a Tuesday.

Mumm is building those five primitives as the substrate for the agent commerce layer. We started with the consumer service — the thing that proves they work end-to-end, in the field, with real money moving and real merchants on the other end. The protocol comes from running that service in production. It is not a thought experiment.

Why this should be open

We could build the protocol layer as a closed walled garden. Several companies in this space are trying to. We think the protocol that wins will be open, and not because of philosophy.

A protocol wins when the cost of integrating it is lower than the cost of building from scratch. Closed protocols carry a lock-in tax that, at scale, gets priced into the integration decision. Open ones don't. So the protocol that wins is the one with the lowest integration cost, and that protocol will be open. We would rather be the most useful implementation of an open standard than the only implementation of a closed one.

We publish our RFCs on GitHub. We commit to backwards compatibility on every primitive endpoint for at least 24 months. Our reference SDKs are Apache-licensed. We will fork ourselves if a competitor does it better, and we expect competitors to fork us when we lead.

What this looks like in three years

Three things will be different in 2029 than today.

First, every consumer-facing transaction over $50 will be mediated by an agent on at least one side. Not most of them — every one. The default behavior of an iPhone, an Android, a Mac, a ChatGPT, a Gemini will be to route the user through a buyer agent before the merchant ever sees the request. The buyer agent will already know the merchant's rate floors and concession policy.

Second, merchants who lack a seller agent will be in the position of merchants who lack a website was in 2005: a small but growing handicap, becoming a fatal one. The CFO who hasn't deployed a Mumm Commerce inbox will be writing apologies to the board about the conversion drop on AI-mediated traffic.

Third, the protocol layer will look like the email layer. Multiple vendors will speak interoperable standards. Most will translate at the edges to support legacy customers. The standards we operate on now — SMTP, HTTP, the credit-card networks — were also the result of someone deciding to publish their RFC and get out of the way.

Why we'd rather be early than right

A lot of agent-commerce thinking right now is brilliantly wrong. Manifestos that imagine economies of agents trading on autonomous decisions, with no human in the loop, no audit trail, no policy band, no escalation. Beautiful, ambitious, and almost certainly never going to ship.

What ships is the boring next step. The buyer agent that calls the hospital billing line on a Tuesday afternoon, gets a $4,000 reduction, hands the user a one-button approval, settles the deal through Stripe, and shows the user a complete transcript. The seller agent that auto-extends a 10% rate concession to a known buyer agent within the bands the revenue manager set last quarter. The audit log that the compliance team can defend in a deposition.

We'd rather be the team that ships the boring next step than the team that wrote the brilliant manifesto. The manifesto is downstream of the boring step working. The boring step ships first.

How this ends well

A future where every consumer transaction is up for negotiation, every consumer has a pro on their side, every merchant has a system that applies policy consistently — that is a future where the average person spends less, the average merchant operates more efficiently, and the friction tax that used to fall on people without time disappears.

It also requires that the protocol underneath be honest. That the agents disclose what they are. That the data they cite is real. That the settlement layer protects both sides. That the reputation layer is grounded in evidence, not vibes. We are trying to build that protocol.

If you're reading this and you want to build on it, build on it. The RFCs are at github.com/trymumm/protocol. The reference SDKs are at github.com/trymumm/mumm-sdk-js and -py. The MCP server is at mcp.trymumm.com.

If you're a merchant reading this and wondering how to handle the wave that's already coming, our Commerce inbox is at trymumm.com/business. The first 100 negotiations a month are free. We expect you to spend the next year deciding whether to build the system yourself or use ours; we'll do our best to win that comparison.

If you're a consumer reading this and you have a bill that's too high, a deal you don't want to do alone, or a perk you suspect was on the table — text us. (910) 555-0100. The first negotiation is free. We don't take a cent unless we save you money.

Mumm is the substrate. The economy on top of it is what we owe the next decade.

“Tell Mumm what you want.”