If you have sat in any AI conversation in the last six months, you will have noticed the same thing we have. The protocol acronyms have stopped being a fringe concern and started showing up in board decks, RFPs, and vendor pitches. MCP, A2A, ACP, AP2, KYA, ATP, AGNTCY. Most people nodding along do not actually know what any of them mean.
This is the reference we wish someone had handed us six months ago. No hype, no vendor framing. Just the acronyms you will keep running into, what they actually do, and which ones matter if you are trying to ship something real.
We will keep it usable. Skim the headings, read the ones you need.
Why this list exists at all
The agentic stack in 2026 is finally past the demo phase. Sierra is at $150M ARR. The MCP registry has crossed nine thousand servers. Gartner is projecting 40% of enterprise apps will have task-specific agents by end of year. Most Fortune 50 companies have an agent in production somewhere.
What changed is not the model quality. It is that the plumbing underneath has started to standardise. Protocols, identity layers, payment rails, discovery mechanisms. The acronyms below are the names of those rails.
If you do not know them, you cannot evaluate a vendor, you cannot brief an engineer, and you cannot tell whether the thing being sold to you is a real interoperability story or a wrapper around a chat box.
The protocols: the part that does the work
MCP: Model Context Protocol
Anthropic’s open standard for letting an agent connect to tools, data sources, APIs, and external systems in a consistent, auditable way. Think of it as USB for agents. The model on one side, your tooling on the other, and a clean spec in between so credentials, schemas, and capabilities are negotiated properly instead of duct-taped.
This is the one to learn first. It has become the default. Google added MCP support to Gemini and Vertex this year, the registry has crossed 9,000+ public servers, and roughly four in five enterprise AI teams have at least one MCP-backed agent in production.
If a vendor cannot tell you how their product speaks MCP, that is a signal.
A2A: Agent-to-Agent Protocol
Originally Google’s, now under the Linux Foundation. A2A lets agents from different vendors and frameworks find each other, hand off tasks, and collaborate. Each agent publishes an Agent Card at /.well-known/agent-card.json so others can discover its capabilities. Transport is JSON-RPC 2.0 over HTTPS.
The practical effect is that a Google ADK agent can invoke a LangGraph agent can invoke a Claude agent, and none of them have to know each other’s internals. This is the closest thing the agent world has to HTTP, and it is the layer where multi-vendor orchestration becomes plausible.
ACP: Agent Communication Protocol
IBM’s parallel effort, originally separate from A2A and now merged with it under the Linux Foundation. When you see ACP referenced in older material, treat it as part of the same converged standard. The fact that the two camps merged is the news, not the spec itself.
AGNTCY: the Internet of Agents collective
Cisco-led industry collective building open standards for cross-vendor agent communication, identity, and discovery. Not a protocol on its own, more a coalition that sponsors and pushes a set of them. Worth knowing the name when it shows up in vendor announcements, because it tells you which side of the standards politics they sit on.
AP2: Agent Payments Protocol
The standard that lets an agent buy something on your behalf without you handing over a card and a prayer. Signed mandates, one-time virtual cards, user-approved spending limits, structured intent. Stripe Link wallets support it. This is the rail that makes autonomous commerce something other than a fraud team’s nightmare.
If you are thinking about agentic checkout, returns, or any agent that touches money, this is the protocol you need to be reading.
UCP: Universal Commerce Protocol
The shopping equivalent of A2A. Lets agents browse catalogs, manage carts, compare prices, and complete checkouts across supported merchants without each merchant having to build a bespoke integration. Often described as HTTP for shopping agents. Pairs naturally with AP2 on the payments side.
KYA: Know Your Agent
The identity and trust layer. Verifies who an agent belongs to, what permissions it has, what it is allowed to do, and crucially, which human or legal entity is on the hook when it does something. Experian, Visa, and Cloudflare are all building implementations. Agent Trust is the current branded version most people are referring to.
If your compliance team is going to let an agent touch production, they will ask about this whether they call it KYA or not.
ATP: Agent Trust Protocol
The IETF-track version of the same idea, recently submitted by Lyrie. Aims to give agent identity and trust the same standards-body treatment that TLS got for encryption. Early days, but worth tracking, because the version that wins becomes the layer every enterprise auth team eventually has to integrate.
The patterns: the part that shapes the design
MAS: Multi-Agent Systems
Architectures where multiple specialised agents, researcher, planner, executor, critic, coordinate on a problem no single agent could finish on its own. This is the design pattern most serious enterprise deployments are converging on, because the single-agent autobot is a 2024 idea that did not scale.
The Astraeus practice is built almost entirely on this pattern. Single agents demo well, multi-agent systems ship work.
HITL: Human-in-the-Loop
A human stays in the workflow for oversight, exception handling, final approval, or ethical intervention. Still the dominant safety approach in production enterprise deployments, and likely to stay that way for anything regulated. Anyone selling you a fully autonomous agent for a critical workflow is selling you a future liability.
RAG: Retrieval-Augmented Generation
The pattern where an agent retrieves relevant documents or context from your data before generating a response, instead of relying purely on what is in the model’s weights. The base technique most “chat with your docs” products are built on, and a core component of almost every serious agent stack underneath the orchestration layer.
LLM: Large Language Model
The model itself. Included for completeness, because half the audience is still using it as a synonym for AI in general and the other half is using it to mean something quite specific.
The platforms and kits: the part that ships
ADK: Agent Development Kit
Google’s four-language SDK (Python, TypeScript, Java, Go) for building agents that natively speak A2A. If your engineering org is already in the Google stack, this is the obvious starting point. The Java and Go releases shipped earlier this year, which matters if you are in a JVM or Go-heavy enterprise environment that has been quietly waiting for someone to take them seriously.
Microsoft Agent Framework
The merger of AutoGen and Semantic Kernel. RC 1.0 shipped February, GA targeted Q1 2026. If you are a Microsoft shop, this is the path. Be aware of the AutoGen / AG2 split. Microsoft is pushing the new AutoGen v0.4+, the community is maintaining AG2 as a fork off the proven v0.2 lineage. Pick deliberately.
LangGraph
Still the highest production-readiness rating in the open-source field, with LangSmith for observability and checkpointing as its real moat. Not an acronym, but you will see it everywhere, and it is worth knowing whether you are looking at a LangGraph implementation or a hand-rolled one.
CrewAI, PydanticAI, Claude Agent SDK
The other names you will see. CrewAI for rapid prototyping. PydanticAI for typed, FastAPI-style agent code, gaining ground in production. Anthropic’s Claude Agent SDK for teams already deep in the Claude ecosystem.
NLWeb
Microsoft’s effort to turn the web into a first-class agent interface, by giving sites a standard way to expose their content and actions to agents. Less mature than the others on this list, but the direction of travel is clear. If your business has a public web presence, this is the layer that will eventually decide whether agents can transact with you or just scrape you.
How to actually use this list
Three honest moves.
If you are evaluating a vendor, ask which of these they implement, which they emit, and which they consume. If they cannot answer cleanly in three sentences, the product is not as far along as the deck suggests.
If you are briefing an engineer or a consultant, name the rails. “We want an MCP-backed agent with A2A handoff to our existing LangGraph orchestration, KYA-compliant identity, and HITL on anything that touches the GL.” That sentence is unambiguous. “We want an AI agent” is not.
If you are setting strategy, decide which two or three of these your business actually depends on, and ignore the rest until they become load-bearing. Most teams need MCP, some form of identity, and a deliberate choice between A2A interoperability and a closed single-vendor stack. The rest is noise until it isn’t.
What we do with all this
Inside Astraeus Business Solutions, we design and run governed multi-agent systems for teams that have outgrown single-agent demos. Which in practice means we make protocol decisions like the ones above on behalf of clients, then build, govern, and operate the resulting stack.
The acronyms are not the point. The point is that the team running the orchestration knows which acronym to reach for, why, and what it costs to be wrong. That is the actual edge.
An offer
Two things you can take us up on.
First, if you are mid-evaluation and want a second opinion on what a vendor is actually shipping under their agent branding, send it over. We will read the deck, the docs, and the protocol claims, and tell you straight.
Second, if you want this list as a living reference inside your own operating system, vault, wiki, Notion, whatever you run on, we can hand you a structured version of it with linked entries, source citations, and the relevant decision questions for each protocol. Reach out and ask.
The agentic plumbing has finally started to standardise. The operators who learn the names of the pipes now are the ones who will still be making sensible calls when the rest of the market is still working out what the letters stand for.
Astraeus Business Solutions is an operator-led AI consulting practice designing governed multi-agent systems for teams that have outgrown single-agent demos. Reach out if you want to talk.