thearm.ai

Reference Implementation · Live in Production · Salt Lake City

The First ARM.

The world’s first working Agent Relationship Manager. Created by Chris Hudson and Ben Reuling. Deployed at a real dealership, Mark Miller Subaru, handling real agent traffic across three open protocols. First deployment April 2026. The first agent-to-agent deal in U.S. history closed through it on May 4, 2026.

The rubric was a definition. This is the definition running.

LIVE
System status
3
Protocols active
334ms
P95 latency
6·2·2
Live · Constrained · Building
MCPModel Context Protocol LIVE
A2AAgent to Agent LIVE
UCPUniversal Commerce Protocol LIVE
AP2Authorize Payments PLANNED
AG-UIAgent-User Interaction PLANNED
A2UIDeclarative Generative UI PLANNED
MCP

Model Context Protocol

The primary protocol for agent-to-tool communication. Agents using MCP connect directly to inventory, pricing, and availability data. The dominant traffic source in the current deployment.

A2A

Agent to Agent

An open protocol for agent-to-agent communication, governed by the Linux Foundation. The channel through which the dealer’s agent and the buyer’s agent speak directly at machine speed.

UCP

Universal Commerce Protocol

An open commerce protocol developed by Google with Shopify, Walmart, Target, and over 20 global partners. Every interaction produces a signed, cryptographically provable record.

In production

Five surfaces. Running today.

Audit logReal-time, across MCP, A2A, and UCP
Live inventory636 vehicles at last count, machine-readable
TrendsThe reservation funnel, visualized
Pricing surfaceFees and disclosures, controlled per rooftop
A2A consoleTest the protocol against the live deployment

This site was a manifesto. Now it is documentation. The next dealer to join will not build any of this from scratch. The work has been done.

Scored against its own standard

Ten criteria. Honest marks.

An ARM that grades itself on a curve is not a standard, it is a brochure. Here is exactly where this implementation stands, and where each claim can be checked.

CRITERION 01

DetectLIVE

Agent traffic is classified at entry by protocol. MCP, A2A, and UCP sessions separate from human sessions before routing begins. Every interaction tagged at ingestion. Hound extends detection to the public web surface.

CRITERION 02

DeployCONSTRAINED

The deployment's transacting agent is in production and meets buyer agents autonomously. Today it transacts at asking price: quotes, availability, 30-minute soft holds, deal handoff. The negotiation tools (submit_offer, counter, accept, reject) are defined in dmc-12 v0.2 and not yet enabled at the reference deployment. When they turn on, this line changes. Not before.

CRITERION 03

EnforceCONSTRAINED

Pricing, doc fees, and tax disclosure are encoded in the schema and held on every quote. No parameter in the public surface accepts a proposed price. That is the strictest enforcement posture there is: the machine cannot be moved off the listed number. Parameter-band enforcement ships with negotiation.

CRITERION 04

RouteLIVE

Agent and human leads travel separate paths from the moment they enter. Human leads route to the CRM. Agent leads route to the ARM. Same dealership, same inventory, two systems.

CRITERION 05

ReportLIVE

Agent activity reports as its own line. Requests, error rate, active agents, latency, throughput, tool mix, and the audit log, live in the dashboard.

CRITERION 06

AdaptBUILDING

The audit trail accumulating now is the training ground. The pattern engine that tunes the deployed agent on outcomes comes online next.

CRITERION 07

VerifyLIVE

Agent identity is verified cryptographically before any tool access. Named agents check against trusted registries. Advocates, scrapers, and probes each get a distinct response. No identity, no transaction.

CRITERION 08

OperateLIVE

A dedicated agent lane runs alongside the dealership stack, drawing on its tools when needed. Humans watch from the dashboard. Discovery and delivery stay human.

CRITERION 09

RememberLIVE

Every interaction carries a cryptographic trace. Identity, protocol, tool, status, latency, trace ID. Persistent, non-repudiable, queryable, 90-day audit trail. When an agent returns, the system knows it.

CRITERION 10

ForecastBUILDING

Forward projection of agent volume and margin impact from accumulated interaction data. The foundation is the audit trail being written today.

Don’t take this page’s word for it. Every live claim maps to a public, unauthenticated document: the Agent Card, the MCP server card, the UCP manifest, the JWKS. Check the well-knowns at dmc-12.ai.

The architecture

Five pillars. One system of record.

01

Agent Registry

Every agent that transacts has an identity, authorized capabilities, and a verifiable key. No identity, no transaction.

02

Trust and Reputation

Every interaction produces signal. Success rate, latency, adherence to terms. Reputation gates access the way credit scores gate lending. The dataset is building in real time.

03

Interaction Logging

Every call, every commitment, recorded with a cryptographic trace. This pillar operates in production today.

04

Protocol Management

The rules of engagement are explicit, versioned, machine-readable. Partners read a schema and plug in. No phone calls. No PDFs.

05

Human Escalation

The conditions that route a transaction to a person are written into the protocol itself. The schema is the policy.

The Definition

Not a feature. Infrastructure.

At its core, The ARM is the ability for agents to interact with each other without friction. Think of it as a highway. We are building the road so the buyer’s agent and the seller’s systems can meet, transact, and move at machine speed.

On that road, an agent can:

SearchLive inventory, every record freshness-stamped
PriceA real, itemized out-the-door price, built from the dealer’s price parameters, taxes and fees included
TradeTrade-in values and payment estimates, every assumption labeled
HoldA hold on a specific VIN, with a status the agent can check
DeliverThe deal into the dealer’s existing systems, in the format the store already consumes

Nothing about the store’s process changes on day one. But the lead that arrives is categorically better than a form fill. It arrives as a nearly finished deal, not a name to chase. Trade structured. Payment estimated. Vehicle held.

The rails carry the deal. The parameters govern it. The dealer set them in advance.

The Doctrine

Five rules.

Rule 01

No Fabricated Numbers

Every price is itemized and computed, never generated. The AI reads tools. It does not improvise a price. Ever.

Rule 02

Consent on Every Handoff

Nothing moves from machine to human without structured consent. The buyer decides when the deal becomes a lead.

Rule 03

The Data Tells the Truth

Freshness is published only when a sync provably succeeded. An unknown odometer stays unknown, never zero. And if a store’s data feed fails, the system flags the failure and holds the last verified state. It never tells an agent the lot is empty when the truth is the feed broke.

Rule 04

The Audit Trail Is the Product

Every conversation is a numbered thread. Every quote, hold, and handoff lives on it, with provenance. One record. Zero disputes about what happened. And the trail compounds: it is the data we learn from to improve every future interaction.

Rule 05

The Parameters Are the Authorization

The dealer sets the parameters in advance. Inside them, the deal runs without stopping, from quote through financing to delivery, as far as the dealer’s tech stack can carry it. Only a deal outside the parameters waits for an answer. Authorization is the constant. Waiting is not.

Trust

It runs both ways.

A deal is a mutual agreement between a buyer and a seller. Agentic commerce does not change that. The ARM does not control the buyer’s agent. It guides the interaction, the way good retail always has: openness first, commitment when both sides are ready.

BrowseAny agent can read inventory and pricing. No sign-up. Open by default.
RegisterA known agent can deliver deals and place short holds. Same trust as a website form, now with a reputation both sides can see.
VerifyWhen a deal needs real commitment, the buyer behind the agent completes a secure, card-backed step. The agent continues. Both sides now know the other is real.
ContractSigned partners get volume, negotiation, and routing.

Reputation is earned in both directions. Good behavior earns reach. Bad behavior earns regulation, applied per agent, not per channel. The ARM gives both sides the ability to manage that. An email inbox never could.

The Standard

What an ARM has to do to be called one.

Before this was a deployment it was a definition. An Agent Relationship Manager is the software layer that orchestrates a company’s relationships with its AI agents, and the relationships between those agents, governing identity, permissions, behavior, and coordination with humans. The definition is only useful if it can be failed, so it comes with ten criteria anyone can score a system against, including this one.

01 — 05Detect · Deploy · Enforce · Route · Report
06 — 10Adapt · Verify · Operate · Remember · Forecast

The full criteria, with what each one means and how to test it, live at thearm.ai/rubric. That is the canonical URL for the standard, and it stays that way — cite it, argue with it, or score your own system against it.

The standard is not automotive. The same ten criteria apply anywhere agents transact: see Industries for how the architecture carries over, and About for who wrote it and why it was published in the open.

The window

Three conditions. All true right now.

Machine-to-machine identity is commodity. The protocol specs are mature enough to publish a manifest and be discovered by any major agent runtime. And the counterparty exists: buyer agents crossed from research to revenue, and one of them closed a deal against this system on May 4, 2026.

Anyone building an ARM can measure their progress against what is already running.


The CRM was the infrastructure of the human lead era. This is the first infrastructure of the agent era.

Official Definition

The layer for the agent era.

Agent Relationship Management (ARM) is the software layer that orchestrates a company’s relationships with AI agents and the relationships between those agents, governing their identity, permissions, behavior, and coordination with humans, the way CRM orchestrates relationships with customers.

CRM orchestrates customers.
ARM orchestrates the agents that orchestrate customers.


“Every dealer eventually had a website. Every dealer will eventually have rails. The category is named. The doctrine is written. The first store is live.”

Chris Hudson · General Manager, Mark Miller Subaru · chris@thearm.ai