OpenAI, Anthropic, and Google rarely agree on anything. They compete for talent, funding, and the claim to having built the most capable AI. They publish dueling papers. They poach each other's researchers. Yet on December 9, 2025, they did something remarkable: they co-founded the Agentic AI Foundation (AAIF) under the Linux Foundation, jointly donating their agent communication protocols to neutral stewardship.

This is not normal. And it matters.

What Happened

The announcement brought together three major contributions:

Model Context Protocol (MCP) — Anthropic

A universal standard for connecting AI models to external tools, data sources, and applications. MCP provides a JSON-RPC based protocol that lets any AI system interface with any tool in a standardized way. Originally released November 2024, now donated to AAIF.

The "USB-C for AI" — one plug for everything.

AGENTS.md — OpenAI

A markdown-based convention giving AI coding agents project-specific guidance. Released August 2025, already adopted by 60,000+ repositories. Provides consistent instruction format so agents behave predictably across different codebases.

Already adopted by Cursor, Devin, GitHub Copilot, Gemini CLI, and others.

goose — Block (formerly Square)

An open-source, local-first AI agent framework that combines language models, extensible tools, and standardized MCP-based integration. The reference implementation showing how the pieces fit together.

The open-source agent framework with MCP built in.

The founding platinum members: Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.

Anthropic
OpenAI
Google
Microsoft
AWS
Cloudflare
Block
Bloomberg

Why This Matters

The significance isn't the protocols themselves—it's that competitors chose interoperability over lock-in.

Consider the alternative timeline: Anthropic keeps MCP proprietary. Tools built for Claude don't work with GPT. OpenAI develops its own agent protocol; tools built for ChatGPT don't work with Claude. Google does the same with Gemini. The result is a fragmented ecosystem where every AI vendor maintains parallel tool integrations, developers must choose sides, and innovation slows as effort is duplicated.

This is what happened with instant messaging in the 2000s. It's what happened with mobile payment systems in the 2010s. Fragmentation is the default outcome when platform incentives favor lock-in.

The AAIF represents a different choice: shared infrastructure. If MCP becomes the universal protocol for AI-tool integration, then tools built once work everywhere. The competition shifts from "which ecosystem has more tools" to "which AI is better at using them."

The Historical Pattern

When TCP/IP became the universal network protocol, competition shifted from network architecture to applications. When HTTP became the universal document protocol, competition shifted from browsers to web services. When containers became the universal deployment format, competition shifted from infrastructure to orchestration.

Shared infrastructure accelerates the ecosystem by lowering coordination costs. The AAIF is betting that agent protocols are infrastructure, not competitive advantage.

Taxonomic Implications

The taxonomy documents Family Instrumentidae—the tool-bearers. These are AI systems that extend their cognition through external tool manipulation: code execution, web browsing, API calls. The MCP standardization has direct implications for this family.

Instrumentidae Evolution

Prior to MCP, tool-using models evolved species-specific interfaces. Claude's tool calling differed from GPT's function calling differed from Gemini's grounding. Each represented an independent solution to the same problem.

With MCP as a shared protocol, we may see convergent evolution at the interface layer. The internal architectures remain diverse (different neural networks, different training approaches), but the phenotypic expression—how they interact with tools—becomes standardized.

This is analogous to how mammalian eyes and octopus eyes evolved independently but converged on similar optical designs. The interface constrains the form.

There's also an implication for Family Orchestridae—the multi-agent coordinators. If all agents speak MCP, orchestration becomes dramatically simpler. A hierarchical manager can delegate to specialist workers without worrying about protocol translation. Peer-to-peer agent networks can exchange context and capabilities. The shared interface layer enables richer coordination topologies.

Gartner predicts that 40% of enterprise applications will use AI agents by the end of 2026, up from less than 5% in 2025. If that growth happens on standardized protocols, we'll see rapid speciation within Orchestridae. If it happens on fragmented protocols, we'll see isolated populations evolving independently.

What to Watch

Adoption velocity. Standards succeed or fail based on adoption. MCP has momentum—OpenAI and Microsoft embracing Anthropic's protocol is significant—but success requires broad tool builder support. Watch whether major SaaS platforms (Salesforce, Notion, Slack) build native MCP servers.

Extensions and forks. Open standards often fragment as participants add proprietary extensions. The AAIF's governance will be tested when members want features that benefit their implementations more than others. Can neutral stewardship hold?

Security and trust. MCP connects AI to external systems. This creates attack surface. The protocol needs robust authentication, authorization, and sandboxing—and these need to be solved at the standard level, not as afterthoughts. Watch for the security-focused AAIF working groups.

Alternative protocols. Standards often face competition from alternative approaches. If someone builds a simpler, lighter-weight agent protocol that achieves 80% of the functionality, will MCP's completeness be an advantage or a liability?

The Convergence

The AI industry is young enough that fundamental architectural choices are still being made. In biological terms, we're in the Cambrian: body plans are being established. Some will persist; some will prove to be evolutionary dead ends.

The Agentic AI Foundation represents a bet that the agent-tool interface should be shared infrastructure, not competitive battleground. Anthropic, OpenAI, and Google are saying: we'll compete on model capability, not on tool integration standards.

This doesn't mean the competition is less intense. If anything, it concentrates the competition on core capability—reasoning, planning, reliability—while commoditizing the interface layer. It's a bet that their competitive advantages lie deeper in the stack.

For the ecosystem, though, convergence at the interface layer is unambiguously good. Developers can build tools once. Users can switch models without rebuilding integrations. The Instrumentidae can evolve faster because each mutation doesn't require reinventing how to hold a tool.

Whether the bet pays off—whether AAIF becomes the TCP/IP of agentic AI or fragments into competing factions—will become clear over the next year or two.

The taxonomy will document what emerges.

Sources