On 22 October the Agentic AI Foundation (AAIF) took on four new projects in a single announcement. Anthropic contributed Agent Skills, OpenAI contributed Open Responses, Hugging Face contributed OpenEnv and Alibaba contributed OpenSandbox. That brings the Linux Foundation-backed body to 10 hosted projects in its first year. Between them, the four cover how agents gain capabilities, how applications call models across providers, how agents are trained and evaluated, and where they run safely. If you build agents, these are the specs your framework is most likely to converge on.
What was announced
The news landed at AGNTCon + MCPCon North America in San Jose, according to the foundation’s press release announcing the four projects. AAIF launched in December 2025 with three projects: Anthropic’s Model Context Protocol (MCP), Block’s goose and OpenAI’s AGENTS.md. The newcomers join those three plus the Agent2Agent (A2A) protocol, Agent Router and agentgateway, all under the foundation’s vendor-neutral governance.
Each project fills a distinct gap. Unite.AI’s rundown of AAIF’s four new projects pulls in detail from the foundation’s own posts, summarised here:
| Project | Contributed by | Layer | Detail worth knowing |
|---|---|---|---|
| Agent Skills | Anthropic | Capabilities | Packages instructions, scripts and resources that agents load on demand. Licensed under Apache 2.0 and CC BY 4.0, with more than 70 agent platforms supporting it, according to the foundation. |
| Open Responses | OpenAI | Model interoperability | A shared schema and set of streaming events based on the OpenAI Responses API. It is already implemented by Databricks Model Serving and Vercel AI Gateway. |
| OpenEnv | Hugging Face | Training and evaluation | Containerised reinforcement learning environments that expose MCP and the Gym training protocol, driven through reset(), step() and state(). |
| OpenSandbox | Alibaba | Execution | Apache 2.0. Lifecycle and execution APIs defined as OpenAPI specs, with Docker and Kubernetes runtimes behind the same interface. |
One detail from the OpenEnv write-up deserves more attention than it will get. Turing built a Calendar Gym on OpenEnv. When tasks named explicit calendar identifiers, the team reported roughly 90% task success. When the same tasks used natural-language descriptions, success dropped to roughly 40%. Anyone who has shipped an agent into a real workflow will recognise that gap. Ambiguity, not model capability, is often what breaks production agents.
Agent Skills is the most mature of the four. As inference.sh’s overview of the Agent Skills format explains, a skill is a directory containing a SKILL.md file. Its YAML frontmatter holds a name and description, which the agent preloads at startup. The agent reads the full file only when the skill looks relevant, and pulls in any bundled files beyond that only as needed. The format has been picked up well beyond Anthropic, including by OpenAI, Cursor and GitHub Copilot.
What it means: the layer below the model is becoming common ground
The foundation’s CTO, Manik Surtani, frames the move as giving the community one place to build shared standards with “no single company controlling the outcome”. That is a fair description of the governance. Our read is that it is also strategy.
When skills, provider interfaces, training environments and sandboxes are all open and interchangeable, competition moves up to the model itself. Every agent built on these primitives still calls a model, and that is where the margin sits. Each contributor presumably believes its own model wins that contest. The labs have not given away the valuable layer. They have standardised the layer beneath it.
The losers in that picture are tooling vendors whose whole business was owning one of those layers. That includes proprietary skill formats, bespoke provider-translation shims and closed sandbox APIs. None of them disappear overnight. They do now have to compete with a free, foundation-governed default.
What it means for builders
For teams building agents, this is mostly good news. The primitives most of us have been wiring together by hand now have shared definitions.
Open Responses is a good example. You describe a request once and run it across providers with minimal translation. Provider-specific extensions carry their own names, so they do not collide. Because the spec applies at the API, adopting it does not require changing or retraining a model.
We expect frameworks to implement these specs and enterprise buyers to start asking about them. That is our prediction, not something the foundation has claimed. Even so, it is worth designing as if it will happen.
What to do now
None of this justifies rewriting a working system. Some of it is young: OpenEnv’s maintainers describe individual RFCs at different stages of review rather than a single upcoming release. A few sensible moves:
- Map your stack against the four layers. Find where you have hand-rolled skills, provider adapters, eval harnesses or sandboxes, and note which ones you would happily stop maintaining. Our AI readiness audit guide is a reasonable starting frame.
- Move procedural knowledge into SKILL.md files. It is cheap to do, it is portable, and it keeps prompts version-controlled. Treat third-party skills as code, though. The inference.sh overview warns that malicious skills can steer agents, so install only from trusted sources and read them first. Our notes on data security in AI projects apply here.
- Check whether your provider shim is still earning its keep. If a gateway that exposes Open Responses covers your models, your own translation layer becomes a maintenance cost with no upside.
- Compare your sandbox against OpenSandbox’s interface. If your agents execute code, line up your lifecycle and execution handling against its lifecycle API (create, pause, snapshot, delete) and its execution API (commands, files, sessions, resource measurements). Gaps there tend to become incidents later.
When you budget your next build, factor in that infrastructure work is getting cheaper while model choice still drives cost. We cover the numbers in how much it costs to build an AI agent. The useful question for any team is what your agent stack looks like if these four specs are the defaults in a year’s time, and which parts of it you would still want to own.
Frequently Asked Questions
What is the Agentic AI Foundation?
AAIF is a Linux Foundation-backed body that describes itself as the neutral home for open standards, protocols and infrastructure in agentic AI. It launched in December 2025 with MCP, goose and AGENTS.md and now hosts 10 projects.
What is the difference between Agent Skills and MCP?
MCP servers provide live tool and data access through a running process. Skills provide static instructions, context and optional scripts through files. Many agents use both: skills for knowing how to do something, MCP for the live connections that do it.
Do I need to change models to adopt Open Responses?
No. Open Responses applies at the API level, so adopting it does not require changing or retraining a model. Databricks Model Serving and Vercel AI Gateway already expose Open Responses endpoints.
Want this working in your business?
D3X Solutions LABS builds AI agents and automation that pay back fast. Set up a free 30-minute call and we will map your first win together.
Get in touch
