AI Agent Crypto 101: The Beginner’s Glossary

The AI agent crypto space has its own language: "MPC custody," "x402 payment flow," and "A2A mandate," making it sound like a different world if you're not already familiar.
This glossary breaks it down. If you want a more comprehensive overview of how these concepts connect, check out the AI agent wallet guide. This glossary focuses solely on definitions.

 

LLM

Large Language Model

The reasoning engine inside an AI agent. An LLM is a neural network trained on massive amounts of text that can understand instructions and generate responses. GPT-4, Claude, and Gemini are all LLMs. The LLM is the brain; the agent is what happens when you give that brain a body and a to-do list.

AI Agent

Autonomous software program

A software program that takes a goal, breaks it into steps, uses tools to complete those steps, and adapts based on results, all without a human approving each action. Unlike a chatbot that just responds, an agent acts. It can browse the web, write code, call APIs, and, with the right setup, spend money on your behalf.

Agentic AI

Category/descriptor

The broader category of AI systems designed to take action, not just generate responses. "Agentic" means the system has some degree of autonomy to pursue goals. An agentic payment is one initiated by such a system, with no human clicking confirm at the time it happens.

Orchestrator Agent

Architecture role

The top-level agent in a multi-agent system. It receives the overall goal from the user, breaks it into sub-tasks, assigns those tasks to specialized sub-agents, and manages the overall budget. Think of it as the project manager. It typically controls the master wallet and distributes funds to sub-agents as needed.

Sub-Agent

Architecture role

A specialized agent that handles one part of a larger workflow, delegated to it by an orchestrator. A research sub-agent might pull data; a payment sub-agent might handle all the onchain transactions. Each sub-agent usually gets its own limited wallet, so a problem in one lane can't drain the whole operation.

DID (Decentralized Identifier)

Identity standard

A machine-readable identity credential that lives on a blockchain instead of in a centralized database. In the AI agent context, a DID gives an agent a persistent, verifiable identity that can be checked across different platforms. This matters for compliance: regulations require that a human can be connected to any agent's financial activity, and DIDs provide the audit trail.

Agent Card

A2A protocol component

A JSON file that acts as a digital business card for an AI agent. It describes what the agent can do, what data formats it accepts, what authentication it requires, and how to contact it. When one agent wants to hire another agent for a task, it reads the Agent Card first to understand the capabilities and pricing.

MCP

Model Context Protocol

An open standard created by Anthropic in late 2024 that gives AI agents a standardized way to connect to external tools, databases, and services. Before MCP, every tool integration required custom engineering. MCP standardizes that connection so any compatible agent can plug into any compatible service. Think USB-C: one universal standard instead of a different cable for everything. Major platforms including OpenAI, Google, and Microsoft adopted MCP in 2025.

MCP Server

MCP component

The service-side component of MCP. An exchange, a database, a wallet provider, or any other platform that wants to be usable by AI agents builds an MCP server. The server exposes the platform's capabilities in a standardized format that any MCP-compatible agent can discover and use.

MCP Client

MCP component

The agent-side component of MCP. The AI agent is the MCP client. It connects to MCP servers, reads what they offer, and calls their tools. The client-server architecture keeps things modular: agents do not need to know how each tool is built internally, just what it can do.

A2A (Agent2Agent)

Communication protocol

An open protocol developed by Google and transferred to the Linux Foundation in June 2025. A2A defines how AI agents from different companies or frameworks discover each other, exchange messages, and coordinate tasks. If MCP is how an agent talks to tools, A2A is how an agent talks to other agents. The protocol includes Agent Cards, task handoffs, and authentication standards.

AP2 (Agent Payments Protocol)

Authorization framework

Developed by Google alongside more than 60 partner organizations, including Coinbase, Mastercard, and PayPal. AP2 is not a payment rail; it is an authorization framework. It defines cryptographically signed "mandates" that prove a human authorized an agent to spend on their behalf. Two types: Cart Mandates (human approves a specific purchase in real time) and Intent Mandates (human authorizes a category of spend upfront; agent acts autonomously later). AP2 has a crypto extension built on A2A and x402.

ACP (Agentic Commerce Protocol)

Checkout standard

Co-developed by OpenAI and Stripe to standardize how AI agents interact with merchant checkout flows. ACP handles the conversation layer: how an agent communicates with a merchant catalog, creates an order, modifies it, or cancels it. It runs primarily on fiat rails through Stripe's infrastructure. ACP debuted in ChatGPT's Instant Checkout feature in early 2026.

MPP (Machine Payments Protocol)

Stripe payment protocol

Stripe's protocol for agent-initiated payments was launched in February 2026 on Base. MPP bridges crypto and fiat: agents can pay in stablecoins, and Stripe converts to whatever the merchant needs. MPP focuses on session-based streaming payments and is the enterprise-friendly version of x402, with compliance features built in.

AI Agent Wallet

A cryptocurrency wallet designed specifically for autonomous software to use. Unlike a standard wallet, which requires a human to approve every transaction, an AI agent wallet operates within preset policy rules set by the human owner. The agent can transact freely inside those rules, but the rules themselves act as hard guardrails. For the full explainer, see the AI agent wallet guide.

Private Key

The secret number that gives whoever holds it full control over a wallet's funds. In an agent wallet, the private key must be carefully controlled because the agent needs signing ability, but unlimited access would be dangerous. See the private key security guide for how to protect yours.

MPC (Multi-Party Computation)

Key management technique

A cryptographic method where a private key is split into multiple shares held by different parties or servers. No single share gives full signing power. The agent can trigger a transaction, but completing it requires cooperation between the multiple keyholders. If one server is compromised, the attacker still cannot sign transactions. This is the standard key management approach in enterprise-grade agent wallet platforms.

Smart Wallet

Wallet type

A wallet that is itself a smart contract deployed on the blockchain. Instead of a single private key controlling everything, the smart contract encodes rules directly onchain. Transactions only execute if they pass the contract's logic: is the amount under the spending cap? Is the destination address on the whitelist? Is the session still active? This enforcement happens at the blockchain level, not just in a software layer that could be bypassed.

Custodial Wallet

Custody model

A wallet where a third party (a platform, exchange, or service provider) holds and manages the private keys on your behalf. You log in with a username and password; they control the actual keys. Convenient, but the platform has custody of your funds. In the agent context, most agentic wallets are custodial: the platform manages the agent's keys on behalf of the human who authorized the agent.

Self-Custody Wallet

Custody model

A wallet where you hold and control your own private keys. No third party has access. A self-custody wallet is the gold standard for protecting significant crypto holdings. The core security principle for anyone using agent wallets: keep your main holdings in self-custody. Only fund agent wallets with amounts the agent actually needs for the task.

Hardware Wallet

Wallet type

A physical device that generates and stores private keys entirely offline. The key never touches the internet. A hardware wallet signs transactions internally and broadcasts only the signed result, so even if your computer is compromised, your key is safe. For the human behind an AI agent setup, a hardware wallet for master holdings is critical. The agent's wallet can be hot; your main holdings should not be.

Policy Engine

Agent security component

The rules layer inside an AI agent wallet that defines what the agent is and is not allowed to do. Typical policies include: maximum spend per transaction, daily spending cap, whitelisted destination addresses, permitted token types, and session expiry. The policy engine is what makes delegating financial authority to an agent safe. Without it, the agent has unconstrained access to whatever funds it holds.

Spending Limit

Policy parameter

A hard cap on how much an agent wallet can spend in a given time window or per transaction. Spending limits are the most basic safety control in agent finance. Setting them conservatively is standard practice: fund the wallet with what the agent needs, set the limit at that amount, and nothing beyond that is at risk.

Whitelist

Policy parameter

A list of approved destination addresses that an agent wallet is permitted to send funds to. Even if an agent is compromised or manipulated into trying to send funds somewhere unexpected, a whitelist prevents it. The transaction simply reverts if the destination is not on the approved list.

x402

Payment protocol

An open payment protocol created by Coinbase in May 2025 that embeds stablecoin payments directly into HTTP web requests. The name comes from the HTTP 402 "Payment Required" status code, which existed in the web standards since the 1990s but was never widely used. x402 finally uses it. When an agent calls a paid API endpoint, the server responds with a 402 and payment instructions. The agent pays in USDC, attaches proof to the next request, and receives the resource. No account. No subscription. No billing page. By early 2026, x402 had processed hundreds of millions of transactions and was live on Base, Solana, Ethereum, Arbitrum, Polygon, and Stellar.

HTTP 402

Web standard

The original "Payment Required" HTTP status code was defined in 1991 and was reserved but never officially used. The x402 protocol revives it as the trigger for machine-to-machine payments. When a server returns a 402 response, it is signaling: "You can have this resource, but you need to pay first." The x402 protocol defines exactly what that payment looks like.

Agentic Payment

Payment type

A financial transaction that is initiated, authorized, and completed by software without a human approving it at the time it happens. Humans set the rules in advance. The agent executes within those rules. Agentic payments differ from traditional automated payments in that the amount, timing, and recipient can vary based on the agent's actions, rather than being fixed on a schedule.

M2M (Machine-to-Machine) Payments

Payment category

Financial transactions between two pieces of software, with no human in the payment loop. An AI agent paying another AI agent for compute time, or paying an API for data, are both M2M payments. This category did not meaningfully exist before crypto. Credit cards cannot handle $0.004 per API call. Stablecoins on Layer 2 networks can.

Micropayment

Payment type

A very small financial transaction, typically fractions of a cent up to a few dollars. The agentic economy runs on micropayments: paying per API call, per kilobyte of data, per second of compute. Traditional payment rails charge 2-3% plus a flat fee, making a $0.01 transaction economically absurd. Stablecoins on Layer 2 networks settle for fractions of a cent, making micropayments viable for the first time.

Pay-Per-Request

Payment model

A billing model where you pay for exactly what you use, at the moment you use it, rather than subscribing in advance. x402 is built around pay-per-request: the agent pays for one API call, receives one result, and that's the transaction. No subscription, no account, no annual commitment. This model is only practical with onchain stablecoin payments.

Agent Payments Protocol (AP2)

Authorization layer

Google's framework for proving that a human authorized an agent's financial action. AP2 introduces "mandates": cryptographically signed documents that record the human's intent. A Cart Mandate means the human approved a specific transaction in real time. An Intent Mandate means the human authorized a category of action in advance. AP2 is the compliance and auditability layer, not the payment execution layer.

Stablecoin

Asset type

A cryptocurrency whose value is pegged to a stable reference asset, most commonly the US dollar. USDC and USDT are the dominant stablecoins. In agent wallets, stablecoins are preferred over volatile assets like ETH or BTC because an agent making autonomous financial decisions cannot do reliable math if the currency's value shifts 10% mid-task. Stablecoin transaction volume reached $33 trillion in 2025, up 72% year over year, with agentic payments cited as a key growth driver.

USDC

Stablecoin

USD Coin is a dollar-pegged stablecoin issued by Circle and managed by the Center consortium. USDC is the dominant settlement token in the AI agent ecosystem. It is the primary asset used in x402 transactions and in Coinbase's Agentic Wallet infrastructure. Each USDC is backed 1:1 by US dollar reserves held in regulated financial institutions.

Stablecoin Rails

Infrastructure concept

The blockchain infrastructure that carries stablecoin transactions from sender to recipient. "Rails" is borrowed from payments terminology ("card rails," "ACH rails"). Stablecoin rails refer to the combination of the blockchain network, the stablecoin smart contract, and the layer of protocols built on top. For AI agents, stablecoin rails replace card rails because they are programmable, open to software without KYC, and settle in seconds.

Gas Fee

Network cost

The fee paid to the blockchain network to process a transaction. Gas fees compensate the validators who confirm and record the transaction. On Ethereum mainnet, gas fees can be dollars; on Layer 2 networks like Base or Solana, they are fractions of a cent. For high-frequency agent micropayments, Layer 2 gas fees are critical to making the economics work.

Layer 2 (L2)

Blockchain infrastructure

A secondary network built on top of a base blockchain (Layer 1, like Ethereum) that processes transactions faster and cheaper. Base (built by Coinbase), Arbitrum, and Polygon are all Layer 2 networks. Most agentic payment activity runs on L2s because the speed and cost profile fits micropayments. x402 launched primarily on Base and Solana for this reason.

Base

Layer 2 network

Coinbase's Layer 2 network built on Ethereum. Base is the home network for x402 and Coinbase's Agentic Wallet infrastructure. Transactions on Base settle in around 200 milliseconds and cost fractions of a cent. Coinbase's Payments MCP recorded a 10,000% spike in agent transactions on Base in early 2026.

Onchain

Location descriptor

Recorded and settled on a blockchain. When a payment is described as "onchain," it means the transaction is recorded permanently and publicly on the blockchain. Onchain transactions cannot be reversed or altered. This is both the security advantage and the risk: agent mistakes are not easily undoable.

Smart Contract

Blockchain program

Self-executing code deployed on a blockchain. A smart contract defines rules that execute automatically when conditions are met, with no intermediary needed. In the AI agent context, smart contracts enforce spending policies in smart wallets, power the DeFi protocols agents interact with, and settle x402 transactions through facilitators.

Facilitator

x402 component

In the x402 protocol, the facilitator is the component that validates payment proofs and submits transactions to the blockchain. When an agent pays for a resource, the facilitator checks that the payment proof is valid before the server delivers the content. Coinbase provides public facilitators for Base and Solana. The facilitator is a centralized point: if it goes down, x402 transactions on that network stop.

Gasless Transaction

Payment UX feature

A transaction where the end user or agent does not need to separately acquire the native network token to pay gas. Instead, gas is abstracted away, either sponsored by the platform or paid in the stablecoin being transferred. Gasless transactions simplify the agent payment experience significantly: agents do not need to maintain ETH, MATIC, or SOL balances just to pay network fees.

Non-Custodial

Custody model

A wallet where you, not a platform, hold the private keys. Non-custodial means no third party can freeze, seize, or lose your funds. For your main crypto holdings, non-custodial is the right model. For an agent wallet, the platform typically holds the keys on your behalf (custodial), which is why funded amounts should be small and spending limits tight.

Cold Storage

Security practice

Keeping private keys completely offline, out of reach of any internet-connected software. A cold storage wallet like Tangem keeps your keys on a hardware chip that never connects to the internet. As the agent economy grows and more software gets payment permissions, cold storage for core holdings becomes more important, not less.

Attack Surface

Security concept

All the points in a system where an attacker could try to gain unauthorized access. Each AI agent wallet you authorize, each permission you grant, and each protocol you plug into adds to your attack surface. Managing attack surface in an agentic setup means giving each agent the minimum permissions it needs, and no more.

Rug Pull

Scam type

A crypto scam where a project's developers abandon it and run off with user funds. In the agent context, a "rogue agent" or "malicious MCP server" is the modern equivalent: software that appears to provide a service but routes agent payments to an attacker's wallet instead. Whitelisted destination addresses in the policy engine are the defense.

KYC (Know Your Customer)

Regulatory requirement

The identity verification process that financial institutions must perform before providing services. Banks, exchanges, and payment processors require government ID to onboard users. AI agents cannot satisfy KYC requirements because they are software. This is one of the core reasons agent payments are built on crypto rails instead of card rails: crypto wallets do not require KYC to operate.
 

Sources

The following sources were used to verify definitions in this glossary:

1. Anthropic: MCP Documentation
2. Google / Linux Foundation: A2A Protocol Specification
3. Coinbase Developer Platform: x402 Protocol
4. MoonPay: Why Agentic Payments Are the Future of AI and Crypto
5. Crossmint: Agentic Payments Protocols Compared
6. Stellar Foundation: x402 on Stellar
7. BeInCrypto: Agentic Finance Explained
8. CCN: AI Agents Payment Rails Compared (MPP, ACP, AP2, x402)
9. Ethereum Foundation: Smart Contract Documentation