Open Responses joins a common foundation for AI agents

OpenAI donates Open Responses to the Agentic AI Foundation, which also hosts Agent Skills, OpenEnv, and OpenSandbox. The goal: reduce fragmentation of APIs and software components for AI agents.

OpenAI has transferred Open Responses to the Agentic AI Foundation, while Anthropic, Hugging Face, and Alibaba also contribute Agent Skills, OpenEnv, and OpenSandbox. This batch of projects, announced on October 22, 2026, brings the foundation to 10 hosted projects and strengthens its role as a neutral gateway for AI agent components. The goal is simple: prevent each provider from imposing its format, tools, and little API quirks.

Why this shift matters for developers

Open Responses, an LLM API specification designed by OpenAI, aims to standardize common primitives between models: messages, tool calls, streaming events, and response formats. In practice, it allows describing a request once and circulating it among multiple providers with less platform-specific translation. OpenAI indicates that Vercel AI Gateway and Databricks Model Serving already implement it, giving initial traction beyond the circle of initial contributors.

The issue goes beyond mere technical elegance. In a market where AI agents must talk to models, execute code, load skills, and rely on training environments, each additional layer can become a friction point. A neutral foundation then serves as a buffer zone between commercial interests of vendors and interoperability needs of product teams.

Four projects, four pieces of the agent puzzle

The Agentic AI Foundation presents the newcomers as covering different layers of the agent stack: Agent Skills to provide additional capabilities to agents, Open Responses for the LLM interface, OpenEnv for reinforcement environments, and OpenSandbox for isolated execution. Together, they cover preparation, execution, and evaluation of agents.

Agent Skills, led by Anthropic, provides an open format to package instructions, scripts, and reusable resources. The foundation says more than 70 agent platforms already support it, with over one million skills published and nearly 280 million installations. OpenEnv is described as an open-source framework for reinforcement learning environments: a client launches an episode, sends actions, and receives observations, rewards, and metadata. Hugging Face states that teams like Turing have already built on it, with very mixed results depending on how tasks are formulated [to be verified on exact test conditions].

OpenSandbox, finally, provides a sandbox component for AI applications, with multi-language SDK, unified APIs, and Docker and Kubernetes runtimes. Its architecture relies on a lifecycle API to create, suspend, resume, or delete a sandbox, and an execution API to run commands and code. Less flashy than a chatty assistant, but much more useful when preventing an agent from doing anything with production keys.

A heavier foundation, but still sellable as neutral

Created in December 2025 under the Linux Foundation’s aegis, the AAIF’s mission is to host standards, protocols, and infrastructures without any single vendor locking the path. At its launch, it had three founding projects: MCP, goose, and AGENTS.md. With the four new ones, it reaches 10 hosted projects, according to the October 22, 2026 announcement.

This consolidation does not erase fault lines. OpenAI pushes its own spec, Anthropic defends its Messages API, Google relies on its Interactions API [to be verified for the exact name according to public offerings], and each keeps its habits. But the presence of all these names in one foundation reflects a more pragmatic reality: AI agents are becoming widespread, and the industry needs common ground to avoid rewriting integrations with every provider change.

What the market is watching now

The real question is not whether these standards are elegant. It is whether they will be widely adopted enough to reduce integration, maintenance, and migration costs. Open Responses targets precisely this pain point: less translation between APIs, fewer vendor-specific integrations, thus less technical debt when a team changes model or provider.

In this context, the AAIF gains credibility by hosting projects from direct competitors. The foundation becomes a place of compromise, sometimes a bit too polished to be entirely innocent, but useful to industrialize agents that must operate in multiple environments without breaking at the first variable change.

Key points

  • AAIF hosts 10 projects since October 22, 2026.
  • Open Responses comes from OpenAI, Agent Skills from Anthropic.
  • More than 70 platforms support Agent Skills.
  • OpenEnv claims 280 million skill installations.
  • AAIF was launched in December 2025 under the Linux Foundation.

By the numbers

  • 10 projects — hosted by AAIF after the October 22, 2026 announcements.
  • 3 projects — founding projects of the foundation in December 2025.
  • More than 70 platforms — compatibility announced with Agent Skills.
  • Nearly 280 million — skill installations mentioned by the foundation.
  • Over one million — skills published on skills.sh in seven months.

To read