CNAI & AAIF#

Sub-chapter 5 of Open Source Projects · The cloud-native AI map and the agentic AI standards body

Two abbreviations that are about to get loud. CNAI (Cloud Native AI) is the CNCF working group defining how AI workloads run on cloud-native infrastructure. AAIF (Agentic AI Foundation) is the brand-new Linux Foundation entity - launched in December 2025 - hosting the standards for the agent layer above that.

If you're entering the field in 2026, this is the chapter that places the current AI infra discourse in your head. These two initiatives are how the open-source community is coordinating on the most consequential platform shift since Kubernetes.

TL;DR#

Mental model: CNAI maps the runtime; AAIF standardises the agent protocols above it.

CNAI - Cloud Native Artificial Intelligence#

What it is#

CNAI is not a single project - it's an umbrella term and a CNCF Working Group (cnaiwg, under TAG Runtime). Its job is to coordinate the patterns, best practices, and reference architectures for running AI/ML workloads on cloud-native infrastructure.

It produced its flagship Cloud Native AI whitepaper at KubeCon Europe in March 2024 (Adel Zaalouk et al.). Follow-up whitepapers in flight cover AI scheduling, AI security, and AI risk / NIST AI Risk Management Framework alignment.

The thesis#

AI/ML workloads - training, fine-tuning, inference, RAG, agents - increasingly run on Kubernetes. But they stress schedulers, networking, storage, and observability in new ways:

• GPU scheduling is harder than CPU scheduling.
• Model serving has latency / batching / autoscaling profiles unlike web traffic.
• Data pipelines for training move terabytes, not request payloads.
• Security has new attack surfaces (prompt injection, model exfiltration).
• Observability must capture model behaviour, not just system metrics.

CNAI documents the gaps and points to CNCF projects that address them.

The CNAI map#

   ┌───────────────────────────────────────────┐
   │ Agentic layer (AAIF - MCP, goose, …)      │
   ├───────────────────────────────────────────┤
   │ Serving (KServe, vLLM, AI Gateway, …)     │
   ├───────────────────────────────────────────┤
   │ Model lifecycle (Kubeflow, MLflow)        │
   ├───────────────────────────────────────────┤
   │ Data (Fluid, Alluxio, object stores)      │
   ├───────────────────────────────────────────┤
   │ Orchestration (Volcano, Kueue)            │
   ├───────────────────────────────────────────┤
   │ Infra (K8s, GPUs, device plugins)         │
   └───────────────────────────────────────────┘

Governance and license#

• Hosted by CNCF (Linux Foundation).
• Lives under TAG Runtime with co-sponsorship from TAG Observability.
• All artifacts under CC-BY-4.0 (docs) / Apache 2.0 (code).
• Meetings every 2nd & 4th Thursday, 10–11am PT.

Get involved#

• Read the whitepaper: cncf.io/reports/cloud-native-artificial-intelligence-whitepaper
• File issues / propose WG topics: github.com/cncf/tag-runtime
• Slack: CNCF Slack #wg-artificial-intelligence (slack.cncf.io)
• Mailing list: lists.cncf.io/g/cncf-tag-runtime
• Charter: cncf.io/wgs/cnaiwg/charter

AAIF - Agentic AI Foundation#

What it is#

The Agentic AI Foundation is a Linux Foundation foundation launched in December 2025 by Anthropic, OpenAI, and Block to host neutral, open standards and infrastructure for agentic AI. Founding Platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.

The thesis mirrors what CNCF did for cloud-native: agree on interoperable protocols early so the ecosystem competes on services rather than lock-in.

The three anchor projects#

• MCP - Model Context Protocol. A JSON-RPC protocol where servers expose tools, resources, and prompts to LLM clients. Sometimes called "the USB-C for AI agents." SDKs in Python, TypeScript, Go, Rust, etc. Now adopted by Claude, ChatGPT, IDEs, and dozens of agent frameworks.
• goose. Block's open-source local-first agent framework. Extensible runtime that uses MCP to plug in capabilities.
• AGENTS.md. A top-level markdown file in a repo that tells coding agents how to work in that codebase - build commands, conventions, constraints. Analogous to README.md but written for machines.

Adoption metrics (Dec 2025 → 4 months in)#

• 170+ member organisations
• 110M+ monthly MCP SDK downloads
• MCP shipped in most major coding IDEs (VS Code, JetBrains, Cursor) and chat clients

Governance and license#

• Linux Foundation foundation.
• Structure: Governing Board, Technical Committee, Working Groups, Ambassador Program.
• Licenses: MCP spec + SDKs MIT; goose Apache 2.0; AGENTS.md convention is CC0-style.

Get started#

# Install goose (Block's local-first agent)
curl -fsSL https://github.com/block/goose/releases/download/stable/download_cli.sh | bash

# Or scaffold a TypeScript MCP server
npm create mcp-server@latest hello-mcp
cd hello-mcp && npm run dev

Tutorial: modelcontextprotocol.io/quickstart.

Get involved#

• AAIF GitHub: github.com/aaif (governance + project-proposals)
• MCP: github.com/modelcontextprotocol - good first issue label
• goose: github.com/block/goose
• AGENTS.md: github.com/openai/agents.md
• Discord: discord.com/invite/9zTwngHAMy

Don't confuse these with…#

The space is noisy. AAIF is not the same as:

• AI Alliance (ai-alliance.org) - an earlier IBM/Meta-led consortium for open AI research and tooling.
• LF AI & Data Foundation - the older LF umbrella that hosts PyTorch, ONNX, Milvus, Feast, MLflow, and others.

All three are real and distinct. AAIF is specifically the agentic standards body; LF AI & Data is the broader AI/ML projects body; AI Alliance is more of a research/policy consortium.

Hands-on Checkpoints#

• Read the CNAI whitepaper (≈40 pages). Note which sections you'd want to deepen in your own work.
• Join CNCF Slack #wg-artificial-intelligence and AAIF Discord. Lurk for a week.
• Run a goose agent locally. Connect it to one MCP server (any).
• Build a tiny MCP server that exposes one tool (e.g., "look up customer by email"). Connect it to Claude or goose.
• Write an AGENTS.md for a personal repo. Test it by running an AI coding session against the repo.
• Watch one CNAI WG meeting recording on YouTube.

Further reading#

CNAI#

• CNCF AI WG (TAG Runtime)
• Cloud Native AI whitepaper
• Whitepaper announcement (Mar 2024)
• CNCF Slack

AAIF#

• AAIF home
• LF launch announcement
• OpenAI AAIF post
• Block AAIF launch post
• Model Context Protocol home
• goose repo

Welzin opinion: Welzin's product surface area lives at the agent layer. AAIF's MCP standard is already foundational to how we ship; expect "is there an MCP for that?" to become a routine question in customer calls. Being early on these standards is a free career accelerant - the people on the founding contributor lists in 2026 will be the senior names in this space by 2028.