In this briefing
- 01Neutral governance now covers four different layers of the agent stack
- 02Skill formats and model requests now form two portable contracts
- 03Training environments and execution sandboxes now use adjacent but distinct interfaces
- 04Google places an open runtime and a cloud preview on one infrastructure path
- →What to watch next
- ↗Sources and verification
- The Agentic AI Foundation accepted four projects at the same time. The four projects respectively cover skill packaging, requests across models, reinforcement-learning environments and execution sandboxes, while the number of projects hosted by the foundation increased from three to ten.
- Anthropic says that more than seventy agent platforms have integrated Agent Skills. Open Responses defines shared request and response structures, semantic streaming events and tool-use patterns. Both specifications still allow differences between implementations.
- Google Cloud announced Agent Compute Engine, which will enter preview in the coming months. Google Cloud also announced the open-source Agent Substrate, which is already available on Google Kubernetes Engine. Google's scale and latency figures come from supplier validation, so enterprises still need to verify those figures in their target environments.
Neutral governance now covers four different layers of the agent stack
On 22 October, the Linux Foundation announced that the Agentic AI Foundation had accepted four new projects. Anthropic contributed Agent Skills, OpenAI contributed Open Responses, Hugging Face contributed OpenEnv, and Alibaba contributed OpenSandbox. The foundation says that the new projects increased the number of projects that it hosts from three when it was established in December (month 12) 2025 to ten.
The four projects address different objects. Agent Skills packages professional knowledge and workflows. Open Responses standardises request and response structures across model providers. OpenEnv connects reinforcement-learning environments to training frameworks. OpenSandbox defines the lifecycle and execution interfaces for agent execution environments. A move into a neutral foundation can change the ownership of project governance, but that move cannot prove that the implementations have reached the same level of compatibility, security or production maturity.
For enterprise adoption, the open foundation is expanding from one connection protocol into a combination of interfaces across several layers. Enterprises may reduce the work required to rebuild skill loaders, model adapters, evaluation environments and sandbox interfaces for every platform. Enterprises still need to accept versions, extension fields, permissions, logs, failure recovery and long-term maintenance responsibilities separately, because shared governance cannot replace operational evidence from a particular implementation.
Skill formats and model requests now form two portable contracts
Anthropic says that more than seventy agent platforms have integrated Agent Skills. Anthropic's public material also says that publishers have released one million skills and users have installed skills nearly two hundred and eighty million times. Each skill contains at least one SKILL.md file, and a skill directory can also contain scripts and reference material. Anthropic presents versioned releases, a formal proposal process and security guidance as next-step proposals, so the current adoption figures cannot prove that those governance mechanisms are already complete.
Open Responses defines shared request and response models, semantic streaming events and tool-use patterns. The specification treats an item as the basic unit for context and tool use, and the specification defines states such as in progress, incomplete and completed for an item. The specification also allows providers to add their own item types, and the specification requires implementers to add an implementer prefix to those types. This extension mechanism preserves product differences, but the same mechanism also means that proprietary types are not portable across providers by default.
For enterprise adoption, the skill format explains how an agent can carry reusable operating knowledge, while the model-request specification explains how an application can describe messages, tool calls and streaming states. Enterprises still need controls for skill provenance, script permissions, version locking and review responsibility. Enterprises also need to test how different model providers interpret standard fields, errors and extension types in practice, because consistent formats cannot automatically produce consistent business semantics.
Training environments and execution sandboxes now use adjacent but distinct interfaces
The Agentic AI Foundation describes OpenEnv as an open framework for creating reinforcement-learning environments. OpenEnv uses containerised services to expose agent interfaces, Model Context Protocol and Gym training protocols, and OpenEnv aims to let different training libraries reuse environments. The foundation describes OpenSandbox as a general-purpose sandbox platform. OpenSandbox provides multilingual software development kits, a unified sandbox interface, and Docker and Kubernetes operating modes.
The two types of project handle different lifecycles. A training environment needs to reproduce states, actions and rewards consistently. An execution sandbox needs to isolate code, commands, files, browser activity or desktop activity. A unified interface can reduce environment-connection code, but a unified interface cannot prove that a training trajectory is reproducible. The interface also cannot prove that a sandbox has prevented data leakage, privilege escalation or access across tenants. The Linux Foundation announcement does not provide compatibility tests across implementations, an independent security assessment or a service level.
For enterprise adoption, evaluation environments and production execution environments can share some container and tool interfaces, but enterprises cannot combine the two environments into one acceptance item. Enterprises need to check dataset versions, randomness, reward definitions, network egress, file persistence, credential injection, resource limits and proof of destruction separately. Open interfaces can provide a replacement path, but deployment configuration and operating records still determine the actual controls.
Google places an open runtime and a cloud preview on one infrastructure path
On 22 October, Google Cloud announced Agent Compute Engine. Google says that the service provides each agent with a hardware-isolated micro virtual machine, and Google says that each micro virtual machine receives a separate identity, short-lived credentials and a persistent disk. Google says that the service will enter customer preview in the coming months. Google also says that some governance capabilities across clouds, hybrid environments and local endpoints will enter preview soon, while fine-grained token and cost controls are planned for preview this quarter.
Google Cloud also announced the open-source Agent Substrate, which organisations can use on Google Kubernetes Engine. Google's validation says that Agent Substrate sustained five hundred activations per second in a test with two hundred thousand agents, while suspend-and-resume latency remained below five hundred milliseconds. Google also says that the architecture can deliver up to ten times the compute density. These figures come from Google's validation material. The public material does not provide the complete workload, hardware configuration, cost, failure distribution or an independently reproduced experiment, so enterprises cannot treat the figures as their own production capacity.
For enterprise adoption, open specifications form an operable foundation only when a runtime provides identity, isolation, persistence, network policy and observability. Google's open-source runtime and managed preview show two delivery routes, but the routes do not have the same availability state, control plane or division of responsibility. Enterprises still need to verify density, recovery latency, security isolation and total cost against the target cluster, the target agent workload and the actual waiting pattern.
What to watch next
- Will the Agentic AI Foundation publish version commitments, compatibility tests, security-response processes and long-term maintenance processes for the four new projects?
- Will different model providers use Open Responses compliance tests to show that basic fields, streaming events, tool calls and error semantics remain consistent?
- Will OpenEnv, OpenSandbox and Agent Substrate provide reproducible tests of isolation, recovery, resource density, cost and failure under public workloads?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing