Three public LinkedIn posts from the last two days, one theme: the boring infrastructure that decides whether agents actually work. Four rival labs donated their agent plumbing to a neutral foundation, a veteran benchmark-skeptic took a knife to Meta's newest model card, and an engineer who builds coding agents explained how the sandbox under them really works. Short quotes, links and what I think. Nothing here comes from anyone's private feed.
The Agentic AI Foundation announced four new hosted projects at AGNTCon + MCPCon in San Jose: Agent Skills (from Anthropic), Open Responses (from OpenAI), OpenEnv (from Hugging Face) and OpenSandbox (from Alibaba), bringing it to 10 projects in its first year - I verified all of it on the foundation's own announcement. Maria Sukhareva's teardown of Meta's latest benchmark table lists mixed sources, mixed harnesses, unreported retries and timeouts, and no confidence intervals, with the only independent composite putting Anthropic's Opus 5.5 ahead 58 to 56; her specific quotes come from her public post, which I read, and I did not independently re-run Meta's numbers. Paul Iusztin's explainer lays out the three real levels of coding-agent sandboxing: OS-level (Seatbelt, Bubblewrap), userspace kernel isolation (gVisor), and microVMs (Firecracker, Cloud Hypervisor). I read all three posts on October 23, 2026.
The foundation's own announcement, dated October 22 from AGNTCon + MCPCon North America in San Jose, says AAIF now hosts 10 projects, up from three launch projects in December 2025. The four new ones map neatly onto the stack. Agent Skills, contributed by Anthropic, is an open format for giving agents new capabilities - packaging instructions, scripts and resources so specialized know-how is portable across agent products; Anthropic says it launched the format a year ago and is donating it so anyone building with skills can shape what comes next. Open Responses, from OpenAI, is an open specification for multi-provider LLM interfaces based on the OpenAI Responses API: a shared schema, consistent streaming events and extensible tooling, so a request written once runs across providers with minimal translation. OpenEnv, from Hugging Face, is a framework and community hub for reinforcement-learning environments - a containerized service exposing an agent interface, MCP, and the Gym training protocol - with a committee that lists Hugging Face, Microsoft, Reflection, Meta-PyTorch, Fleet, Mercor, RadixArk, Modal and Prime Intellect among its users. OpenSandbox, from Alibaba, is a general-purpose sandbox platform with multi-language SDKs, unified APIs and Docker and Kubernetes runtimes for coding agents, GUI agents, evaluation and RL training. They join MCP, goose, AGENTS.md, Agent2Agent, Agent Router and agentgateway under what AAIF CTO Manik Surtani calls open, vendor-neutral governance. My take: a year ago the agent-interop story was mostly MCP; as of this week the neutral layer covers skills, APIs, training environments and execution sandboxes too, and the contributors are the very labs competing hardest with each other. That is how plumbing standardizes.
Maria Sukhareva, whose anti-hype benchmark critiques are a genre of their own, went through the documentation behind Meta's latest model card - 13 pages of tables, she says - and posted what she found, amplified by contract CTO Eric Fraser. Her list, quoted from the post: mixed sources, with some rows from Artificial Analysis, Vals, Surge, Datacurve or Abundant AI, some from Meta's own "best-effort" runs of competitor models, and some Claude numbers lifted from Anthropic's system cards. Mixed harnesses: on SWE-Marathon, GPT-6 Astra ran in Codex and Fable 5.1 in Claude Code - a row Meta itself flags as not a controlled comparison - while DeepSWE mixes reporters without any flag. On OSWorld, Meta ran most models in its own framework while the Opus 5.5 score is Anthropic-reported, with the OSWorld release unspecified. Artificial Analysis runs Opus 5.5 with a default fallback to an older Claude model, and how often that fired is not reported. Retries and timeouts are not reported for any row, which matters on long agentic tasks. No confidence intervals or trial counts, with several gaps under a point. Two rows are Meta's own: MemoryBench is internal, and PSI is a Meta-assembled index with no published weighting. And her kicker: the chart's only independent composite, the Artificial Analysis Index, puts Opus 5.5 first, 58 to 56. Her conclusion is blunt: "I learned nothing about how the model will perform on my tasks as compared with other models." Fraser's framing is blunter still. My take: none of these complaints is new, and that is the point - reporting standards across nearly all providers have not moved, so a public, specific, row-by-row audit with the actual flags quoted is the most useful kind of peer review a benchmark table can get. One caveat from me: I read her post but did not re-derive any numbers from Meta's card, so treat each specific claim as hers, quoted, not as my measurement.
Paul Iusztin, who writes for Manning and built his own coding agent called Decode, posted a compact explainer on what actually stops an agent from touching files outside your project. Three levels, from his post: OS-level sandboxing - Seatbelt on macOS and Bubblewrap on Linux - restricts resource access while sharing the host kernel. Userspace kernel isolation - gVisor - intercepts system calls to add isolation between the agent and the host kernel. And microVMs - Firecracker and Cloud Hypervisor - give a separate guest kernel, the closest option to a traditional VM. His practical note is the part people get wrong: you do not sandbox the whole agent. The harness stays local, and the agent's tools - Bash, Read, Write, Edit - execute inside one shared sandbox so every tool sees the same filesystem. In Decode he ended up with one executor interface behind local execution, Docker and remote Modal sandboxes, so each session gets its own isolated environment. My take: the harness-local, tools-sandboxed split matches how the serious coding agents are actually built, and it is the right mental model for anyone wiring one up this weekend - the agent's brain roams free, its hands work in a glovebox.