Agentic AI Foundation: OpenAI & Anthropic Unite on Standards

On December 9, 2025, something unusual happened in the AI industry: OpenAI and Anthropic—two companies that have spent years competing for market share, talent, and benchmark victories—announced they’re working together. The Agentic AI Foundation (AAIF), launched under the Linux Foundation, unites these bitter rivals alongside Block, AWS, Google, Microsoft, Bloomberg, and Cloudflare around a singular goal: standardizing how AI agents connect to tools, systems, and each other.

This isn’t just another AI consortium announcing vague intentions. The foundation arrives with three battle-tested open-source projects already in hand: Anthropic’s Model Context Protocol (MCP) with 10,000+ servers and 97M+ monthly SDK downloads, OpenAI’s AGENTS.md specification adopted by 60,000+ repositories, and Block’s Goose framework that’s saving their developers 50-75% of their time.

But before we get swept up in the vision of harmonious AI agents working seamlessly together, let’s talk about the reality check. Carnegie Mellon research shows that even the best-performing agents fail 70% of real-world office tasks. Gartner predicts 40% of agentic AI projects will be cancelled by 2027. And while 35% of enterprises claim they’ve “adopted” agentic AI, only 6.5% report actual production deployment.

So what’s really happening here?

The Fragmentation Problem the Agentic AI Foundation Solves

Right now, building AI agents is like constructing a building where every vendor uses different electrical outlets, incompatible plumbing standards, and walls that won’t attach to each other. You can build impressive individual rooms, but connecting them into a functional structure requires custom adapters, security workarounds, and constant maintenance.

“We are seeing AI enter a new phase, as conversational systems shift to autonomous agents that can work together,” said Jim Zemlin, Executive Director of the Linux Foundation. The problem is that these agents currently exist as disconnected islands—each with its own framework, APIs, and integration requirements.

The business impact isn’t theoretical. When agents can’t communicate, organizations face lost context during handoffs, conflicting decisions from separate systems, data silos with incompatible formats, and what industry analysts call “agent sprawl”—duplication of effort across the organization because nothing talks to anything else without custom integration work.

This is where the AAIF’s three founding projects become interesting. They’re not starting from scratch—they’re standardizing protocols that already have significant adoption.

The Three Pillars: MCP, AGENTS.md, and Goose

Think of MCP as “USB-C for AI applications.” It’s an open protocol that standardizes how AI systems integrate with external tools, databases, and APIs. Since Anthropic released it in November 2024, it’s become the de facto standard—OpenAI adopted it in March 2025, Google DeepMind in April, and Microsoft integrated it into Azure AI and Windows 11.

The architecture is deliberately simple: a client-server model built on JSON-RPC 2.0 that lets AI models (clients) call servers exposing tools, resources, and prompts. It’s essentially what the Language Server Protocol did for code editors, applied to AI integrations. The 97 million monthly SDK downloads suggest developers find this approach more appealing than building yet another custom integration.

AGENTS.md takes a different angle—it’s project-specific guidance for AI coding agents. Think of it as “a README for agents.” Released by OpenAI in August 2025, it gives AI assistants like Claude Code a consistent place to find build steps, testing conventions, and project context that would clutter a human-focused README. The fact that 60,000+ repositories adopted it in four months—including 88 AGENTS.md files in OpenAI’s own main repository—suggests it solved a real problem.

Then there’s Goose, Block’s open-source AI agent framework. This is where the rubber meets the road: Block reports that 60% of their workforce uses Goose weekly, and it saves developers 50-75% of their time on certain tasks. It’s built on MCP, making it a reference implementation for how these standards work in practice.

“The technology that will define the next decade can either remain closed and proprietary for the benefit of few, or be driven by open standards for the benefit of all,” said Manik Surtani, Head of Open Source at Block. That’s the pitch, anyway.

Why Competitors Are Suddenly Collaborating

Let’s be honest—OpenAI and Anthropic aren’t forming this foundation out of pure altruism. Both companies have spent considerable engineering resources building agent infrastructure. OpenAI’s GPT-5.1 adaptive reasoning and Anthropic’s extended thinking mode both bet heavily on agents as the future of AI interaction.

The strategic calculation is straightforward: better to collaborate on the plumbing than compete on it. If everyone adopts MCP, Anthropic’s protocol becomes critical infrastructure—but the neutral governance of the Linux Foundation prevents them from weaponizing that position. OpenAI gets the same deal with AGENTS.md. Both companies can then compete on what actually differentiates their models: reasoning quality, speed, cost, and reliability.

This mirrors what happened with Kubernetes. Google donated it to the Cloud Native Computing Foundation (another Linux Foundation project) in 2015, and it became the standard orchestration layer. Google didn’t dominate Kubernetes, but they benefited enormously from having a standardized platform where their cloud services could compete on merit rather than lock-in.

The platinum membership list reads like a who’s who of cloud infrastructure: AWS, Google, Microsoft, Bloomberg, Cloudflare. These companies aren’t participating out of charity—they’re making $350,000 bets that standardized agent protocols will expand the total addressable market faster than proprietary approaches.

The Reality Gap: Hype vs. Production

Here’s where the narrative gets complicated. While 78% of C-suite executives say agentic AI requires a new operating model and 35% claim adoption, nearly 80% report no material earnings impact yet. Only 6.5% have reached production deployment.

That’s not a typo. The gap between “we’re doing agentic AI” (35%) and “we have it in production” (6.5%) is 28.5 percentage points. Most organizations are stuck in pilot purgatory—running experiments that generate impressive demos but never make it past proof-of-concept.

Carnegie Mellon’s research helps explain why. When researchers tested Gemini 2.5 Pro (currently the best-performing agent on benchmarks) against real-world office tasks, it failed 70% of the time. Agents struggle with nuance, unexpected conditions, and the kind of judgment calls that humans handle without thinking.

Gartner’s prediction that 40% of agentic AI projects will be cancelled by 2027 isn’t pessimism—it’s pattern recognition. The cited reasons are escalating costs, unclear value proposition, and inadequate risk controls. Harvard Business Review warns that “organizations aren’t ready for the risks of agentic AI” as the complexity of the risk landscape “ramps up sharply” with autonomous systems.

Security Concerns and Missing Players

MCP’s rapid adoption has come with growing pains. In July 2025, researchers disclosed CVE-2025-53110 and CVE-2025-6514, both remote code execution vulnerabilities. A security audit found nearly 2,000 MCP servers lacking authentication, and researchers demonstrated how combining seemingly safe tools could exfiltrate files.

Anthropic responded with an authorization spec update in June 2025, classifying MCP servers as OAuth Resource Servers and requiring clients to implement Resource Indicators (RFC 8707). But security is an ongoing battle, not a solved problem.

Also notable: who’s not at the table. Apple and Meta—two of the largest AI players—are conspicuously absent from the membership list. Apple’s absence isn’t surprising given their preference for proprietary ecosystems. Meta’s absence is more curious, especially given their commitment to open-source AI with Llama models.

The Standards War Question

The AAIF isn’t the only game in town. Google donated the Agent2Agent (A2A) protocol to the Linux Foundation in June 2025, backed by 100+ companies. The official line is that A2A and MCP are complementary—A2A handles agent-to-agent communication (horizontal), while MCP handles agent-to-tool communication (vertical).

That’s the diplomatic version. Some analysts warn that “in practice there may be a tug of war” and that “most of the churn around standards is more about gaining mindshare and securing business development budgets than solving architecture issues.”

We’ve seen this movie before. Remember the browser wars? The USB vs. FireWire debate? Sometimes the best technical solution wins, sometimes network effects dominate, and sometimes we end up supporting multiple standards indefinitely.

What Happens Next

The AAIF has momentum that previous AI consortiums lacked. Unlike vaporware announcements of future collaboration, they’re standardizing protocols that already work at scale. The Linux Foundation’s governance model has decades of credibility. And the founding companies have put actual code where their press releases are.

But standardization doesn’t guarantee success. The technology still needs to prove it can deliver value that justifies the complexity and risk. That 70% failure rate needs to come down. The 6.5% production deployment rate needs to go up. And the security concerns need solutions, not just spec updates.

Mike Krieger, Anthropic’s Chief Product Officer, said that “donating MCP to the Linux Foundation ensures it stays open, neutral, and community-driven as it becomes critical infrastructure for AI.” That’s the aspiration—MCP and AGENTS.md becoming as fundamental to AI development as HTTP is to the web.

Whether we get there depends less on the quality of the standards and more on whether agentic AI can bridge the gap between impressive demos and reliable production systems. The foundation provides the infrastructure. Now the technology needs to prove it’s ready to build on it.

The Agentic AI Foundation launched with more credibility and adoption than any previous AI standardization effort. That doesn’t guarantee success—but it does mean we’ll find out whether this technology is ready for prime time faster than we would have otherwise. And in an industry moving this quickly, that might be the most valuable contribution of all.