Four rival AI companies gave agent infrastructure to one foundation. Check which layers are portable

If your company had to move its AI agents to another vendor next quarter, which parts would come with you?

Until recently the honest answer was: the prompts, if someone saved them. On October 22 the Agentic AI Foundation, the Linux Foundation project that already hosts the Model Context Protocol, took in four more pieces of agent infrastructure, each contributed by a company that competes with the others:

• Agent Skills, from Anthropic: an open format for packaging instructions, scripts and resources that an agent loads when a task needs them.
• Open Responses, from OpenAI: an open specification for talking to models across providers, based on OpenAI’s Responses API, with a shared schema, consistent streaming events and extensible tooling.
• OpenEnv, from Hugging Face: a framework and hub for reinforcement learning environments, run as containers that expose MCP and the Gym training protocol.
• OpenSandbox, from Alibaba: a sandbox platform for coding agents, GUI agents, evaluation and code execution, with a protocol that defines how a sandbox is created, run and torn down, on Docker or Kubernetes.

The foundation now hosts ten projects, up from three at its launch in December 2025. None of the four is new technology. What is new is who controls them.

Read it as a map of the stack

The useful way to read the announcement is by layer. Each project standardizes one interface between parts of an enterprise agent, and together with the earlier projects they cover a surprising share of it.

Where the Agentic AI Foundation now has a project

Capabilities and instructions

• Agent Skills
• AGENTS.md

What the agent knows how to do

Model interface

• Open Responses

How the application calls a model

Tools and other agents

• MCP
• A2A
• agentgateway
• Agent Router

Execution

• OpenSandbox

Where generated code runs

Training and evaluation

• OpenEnv

Not covered

• Memory and state
• Agent identity
• Evaluation data
• Business policy

Where switching costs accumulate

The bottom row matters most. Every project the foundation hosts describes how components talk to each other. None of them describes what an agent has accumulated: the memory it built about your customers, the skills it wrote for itself, the identity it acts under, the evaluation history that justified its approval. We have argued that lock-in is moving to that accumulated state as models become interchangeable. A standard interface makes it easier to swap the parts around that state. It does not move the state.

Neutral governance is not the same as interoperability

Three cautions keep this from being an easy win.

A spec is not an implementation. Two vendors can both claim support for Open Responses and still differ in which events they emit, which tools they allow and how they handle errors. MCP shows the pattern: it standardized how agents reach tools, and enterprises still found that connectivity is not governance. Ask for conformance tests, not logos.

A neutral home does not erase an origin. Open Responses is based on OpenAI’s API. That makes it practical, because many teams already write to that shape, and it means the first design choices were made by one vendor. Neutral governance gives other companies a vote on the next version, not on the first one.

Portable packaging does not make content safe. Agent Skills makes a skill easy to move between products. Whether the procedure inside it is correct, approved and current is a separate question, and in regulated work a skill is a procedure that belongs under document control. A format that travels well also lets an unreviewed procedure travel well.

The sandbox layer shows the split

The same day, AWS took Lambda MicroVMs, its managed service for running user or AI-generated code in isolated virtual machines, to four more regions, including Spain, for 14 in total. Put next to OpenSandbox, it shows the choice enterprises now face at the execution layer: an open protocol you operate, or a managed primitive a cloud operates for you. Both are reasonable. What matters is that the lifecycle (create, run, inspect, destroy) is expressed in a way your platform team can move, because sandboxes are where generated code meets your network.

What to ask vendors now

When an agent platform appears on your shortlist, add four questions to the evaluation:

1. Which of these specifications do you implement, at which version, and do you pass the project’s conformance tests where they exist?
2. Can we export the agent’s memory, skills and configuration in a documented format, and how long does it take?
3. If we bring our own sandbox or our own model endpoint through these interfaces, which of your controls stop working?
4. Who on your side participates in the foundation’s working groups for the layers we depend on?

The answers will not remove lock-in. They will tell you where it lives, which is the part of the decision you can still plan.