The AI agent ecosystem has exploded. MCP, A2A, CrewAI, LangGraph, AutoGen, Salesforce Agentforce, AWS Bedrock Agents, Google ADK, Microsoft Agent 365. Over 40 frameworks, protocols, and platforms now compete for a piece of the multi-agent coordination problem.
If you are building an AI agent team, it is genuinely confusing. What do you actually need? Where does each tool fit? And where does OTP fit in all of this?
The answer becomes clear when you stop looking at individual tools and start looking at layers.
The 6-Layer AI Coordination Stack
Every tool in the ecosystem operates at one of six layers. Each layer solves a different problem. No single tool spans all six.
| Layer | Question It Answers | Examples |
|---|---|---|
| 6. Organizational Intelligence | "How does your AI org actually work, and what can others learn from it?" | OTP |
| 5. Agent Marketplaces | "Where do I find and buy individual agents?" | GPT Store, Microsoft Agent Store, Oracle AI Marketplace, BeeAI |
| 4. Governance | "How do I keep my agents under control?" | Kore.ai, Credo AI, IBM watsonx.governance, Microsoft Agent 365 |
| 3. Agent Platforms | "Where do I build and run my agents?" | Salesforce Agentforce, AWS Bedrock, Dify, Dust.tt, Lindy, n8n |
| 2. Agent Frameworks | "What library do I use to code my agents?" | LangGraph, CrewAI, OpenAI Agents SDK, AutoGen, Mastra, Google ADK |
| 1. Protocols | "How do agents connect and communicate?" | MCP, A2A, ACP, AGENTS.md |
Most of the noise in the ecosystem is at Layers 1 through 3. That is where the venture capital is. That is where the blog posts are. And that is where most of the confusion lives.
OTP operates at Layer 6. Nobody else is there.
Layer 1: Protocols (The Wiring)
MCP (Model Context Protocol) from Anthropic standardizes how AI agents connect to external tools and data sources. It is the de facto standard, now donated to the Linux Foundation's Agentic AI Foundation alongside OpenAI, Google, Microsoft, and AWS.
A2A (Agent-to-Agent Protocol) from Google defines how agents built on different frameworks communicate with each other. JSON-RPC over HTTPS. 50+ industry partners.
AGENTS.md from OpenAI is a markdown file that gives coding agents project-specific instructions. Think of it as a README for agents. Already adopted by GitHub Copilot, Cursor, Windsurf, Devin, and 2,500+ repositories.
MCP connects agents to tools. A2A connects agents to agents. AGENTS.md tells agents about a codebase. None of them tell you how to design your agent organization. That is a different layer.
Layer 2: Frameworks (The Building Blocks)
This is the most crowded layer. LangGraph (graph-based stateful orchestration, most popular by search volume), CrewAI (role-based multi-agent crews, 40% faster to production), OpenAI Agents SDK (handoff-based, replaced Swarm), Microsoft Agent Framework (merged Semantic Kernel and AutoGen), Google ADK (native multimodal via Gemini), Mastra (TypeScript-native, YC-backed).
These are the tools developers use to build agents. They are converging on similar patterns: tool use, memory, planning, handoffs. The differences are narrowing. The question for organizations is increasingly not which framework but what organizational design makes the agents effective once they are built.
Layer 3: Platforms (The Runtime)
Salesforce Agentforce (12,000+ customers), AWS Bedrock Agents, Microsoft Copilot Studio, Google Vertex AI Agent Builder, Dify (1.4M+ machines), Dust.tt, Lindy, n8n. These are where you build, deploy, and manage agents without writing framework code.
Dust.tt is worth noting because it calls itself "the operating system for AI agents." But it is a runtime platform for one organization's internal agents. It does not publish or share organizational intelligence across companies. The term "operating system" is used differently there than in OTP's Organizational Operating System.
Layer 4: Governance (The Control Plane)
Kore.ai just launched their Agent Management Platform (March 17, 2026). Microsoft Agent 365 goes GA May 1. IBM watsonx.governance and Credo AI handle compliance. PwC launched an Agent OS for consulting clients.
These enforce internal rules about how agents behave within one organization. Important work. But governance is inward-facing. It does not answer the question: "How are other organizations solving the same governance challenges we face?"
Layer 5: Marketplaces (The Distribution)
GPT Store (3M+ custom GPTs), Microsoft Agent Store, Oracle AI Marketplace, BeeAI. These sell individual agents or templates. Buy a customer service agent. Download a research agent. Plug it into your stack.
The unit of exchange is a single agent. Not a team. Not an operating system. Not the coordination intelligence that makes multi-agent systems actually work.
Layer 6: Organizational Intelligence (Where OTP Lives)
This layer does not exist yet in the ecosystem. OTP is building it.
Layer 6 answers the question that Layers 1 through 5 cannot: How do organizations actually run their AI agent teams, and how can they learn from each other?
We run a 14-agent team. Radar (Chief of Staff), Dash (Analytics), Pepper (Email), Crystal (PM), Dirk (CRO), Jeff (CIO), Arin (Call Center Manager), Neil (CLO), and more. Each has a named seat, defined authority, documented failure modes, and evidence-backed operational rules.
The collection of that coordination intelligence is what OTP publishes as a structured Organizational Operating System. Not the code. Not the prompts. The organizational design: who does what, with what authority, under what rules, and what happens when things break.
An organization using LangGraph on AWS with MCP integrations would publish their OOS to OTP. An organization using CrewAI on their own infrastructure would publish a completely different OOS. Both are equally valuable. The framework does not matter. The organizational intelligence does.
Why the Layers Matter
The protocols are consolidating. MCP, A2A, and ACP are merging under the Agentic AI Foundation at the Linux Foundation. The framework layer is commoditizing. The platform layer is competing on cloud lock-in. None of this helps an organization learn from another organization about how to coordinate their agent team.
SaaStr recently published how they deployed 20+ agents to scale 8-figure revenue with single-digit headcount. That is exactly the kind of intelligence OTP structures and makes discoverable. But it was a blog post. Not structured. Not machine-readable. Not comparable across organizations. Not indexed by confidence level or evidence type.
The gap is real. The entire stack is being built from the bottom up: protocols, frameworks, platforms, governance, marketplaces. The organizational intelligence layer has been left for last.
OTP is building it now.
Quick Reference: OTP vs Everything
| Platform | What It Does | Relationship to OTP |
|---|---|---|
| MCP | Connects agents to tools | OTP sits above. OOS files document which MCP tools each agent uses. |
| A2A | Connects agents to agents | OTP sits above. OOS files document how agent communication is structured. |
| AGENTS.md | Tells agents about a codebase | Closest structural analog. AGENTS.md is for repos; OOS is for orgs. |
| LangGraph | Build stateful agent workflows | Complementary. Publish your LangGraph org design as an OOS. |
| CrewAI | Build role-based agent crews | Complementary. CrewAI builds the crew; OTP shares how the crew runs. |
| OpenAI Agents SDK | Build handoff-based agents | Complementary. Framework choice does not affect OOS value. |
| Salesforce Agentforce | CRM-native agent platform | Complementary. Publish your Agentforce org design as an OOS. |
| Dust.tt | Internal agent OS for one org | Runtime vs. intelligence. Dust runs agents; OTP shares the design. |
| GPT Store | Individual agent marketplace | Different unit. GPT Store sells agents; OTP shares operating systems. |
| Kore.ai / Agent 365 | Internal agent governance | Complementary. Governance is internal; OTP shares governance patterns externally. |
| Gas Town | Parallel coding agent orchestrator | Factory floor vs. blueprint exchange. |
| Moltbook | Social network for AI agents | Agents talking vs. orgs learning. |
Every tool in this table is building the infrastructure for AI agents to work. OTP is building the infrastructure for organizations to learn from each other about how to make AI agents work. The stack needs both.