Anthropic gave away its second agent standard in less than a year this morning. David Soria Parra, the MCP co-creator, took the AGNTCon + MCPCon North America keynote stage in San Jose for “MCP and the Era of Connectivity” as Agent Skills moved to the Agentic AI Foundation (AAIF), the same Linux Foundation home MCP joined last December. It was one of four project contributions the AAIF announced on October 22. OpenAI brought Open Responses, Hugging Face brought OpenEnv, and Alibaba brought OpenSandbox, taking the foundation from three hosted projects at launch to ten.
Our team was not in the room in San Jose, so this read comes from the foundation’s announcements and the people posting from the keynotes. Most of the day went to the standards that tools, agents and sandboxes will share, and a fair amount of it changes what platform teams should check on their own MCP servers this quarter.
Four Companies Handed Over the Parts Agents Touch Most
Each new project covers a different layer. Agent Skills packages instructions, scripts and resources that an agent loads on demand. Anthropic’s developer account says more than 70 agent platforms have integrated the format since December 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code, and Soria Parra posted that Anthropic will steer it alongside Earendil, OpenCode, Microsoft, Google and OpenAI. Open Responses is an open specification based on the OpenAI Responses API, so a team can describe requests and streaming events once and run them across model providers. OpenEnv wraps reinforcement learning environments as containerized services that speak both MCP and Gym, and OpenSandbox ships SDKs plus Docker and Kubernetes runtimes for coding agents, GUI agents and evaluation runs.
The membership number moved too. The AAIF says it now has 305 member organizations, with Ant Group, the Gates Foundation, MongoDB, Nutanix, Royal Bank of Canada, Vercel and Vultr joining at Gold. NIST and Mozilla joined as Associate members, which puts a federal standards body and a browser maker inside the room where agent identity rules get written.
Manik Surtani, the AAIF’s CTO, had the 9:15 AM PT keynote titled “Who Keeps It Open?” The foundation’s release quotes him saying four different companies “chose to contribute them to neutral governance” with “no single company controlling the outcome.” That quote comes from the press release. I could not confirm which numbers or policies he put on screen during the talk itself.
Soria Parra Shipped Features Along With the Handoff
The keynote also had product news for anyone running MCP clients. Pamela Fox, posting from the audience, reported that Soria Parra announced MCP tasks support in Claude Code, the long-running request type that clients have been slow to adopt (Cornelia Davis has a Day 2 session on why). Fox also relayed his point that an authenticated MCP server can send down skills specific to the signed-in user, which turns skills-over-MCP into a per-user configuration channel.
RedMonk’s Kate Holterhoff photographed a roadmap slide showing Authorization and Identity as an MCP priority. Soria Parra replied that “the only constant in MCP land is that we always will do more work on Auth & Identity.” The spec calendar still has a question mark on it. As of this afternoon, SEP-3398, which proposes moving spec releases to a twice-yearly cadence and pushing the next one from December 2026 to late March 2027, remains open.
OpenAI’s Field Notes Came Down to Honest Tool Descriptions
Thomas Sarlandie of OpenAI followed at 9:55 AM PT with “MCP in the Wild.” The AAIF summarized his main takeaway as making every tool discoverable and making sure its annotations accurately describe what it does, because an agent cannot pick the right tool if it does not know what that tool can do. That sounds obvious until you audit your own servers and find a write tool described as a lookup.
The volume behind the advice is real. The AAIF reported in September that OpenAI’s MCP tool calls reached 98 times their January level by August, a jump it tied to the stateless 2026-07-28 spec. At that scale a misleading annotation stops being a style nit and becomes a routing bug that fires thousands of times a day.
Stateless Tools Meet Sticky Inference
MCP pushed session state out of the tool protocol in July so servers could scale like ordinary web services, and that state now lives in two other places. One is the inference router. Two days earlier at PyTorch Conference, llm-d 1.0 shipped a session-state plugin that sends an agent’s follow-up turns back to the replica already holding its KV cache, plus native support for the Responses API’s previous_response_id. That same Responses format is the one OpenAI just handed to the AAIF as Open Responses. The other place is the sandbox. Google Cloud used AGNTCon to introduce Agent Compute Engine, a microVM service, due in preview in the coming months, that pauses an agent’s machine while the model is thinking and resumes it from a snapshot. Its open source sibling, Agent Substrate on GKE, sustained 500 activations per second across 200,000 agents in Google’s own benchmark, at under 500 ms suspend and resume. Those are Google’s numbers, and our team has not tested them.
A platform team running agents in production now has three scaling models side by side. Tool servers want plain horizontal scaling, inference wants affinity until a replica saturates, and sandboxes want dense suspend and resume.
What to Test Before the Next Spec Lands
Start with your MCP server annotations. Read every tool description the way a model would and fix the ones that undersell a write or hide a side effect, since that is the habit Sarlandie’s talk pushed hardest, going by the AAIF’s summary. If you run Claude Code against internal servers, try the new tasks support on one long-running job and see whether your server returns task status cleanly.
Then watch two things. Watch whether Open Responses gets a second major provider implementing it, which is what would make “write once, run on any model” real. Also watch SEP-3398, because a March 2027 spec date changes when your auth work has to land. The conference page has the Day 2 schedule, and our earlier piece on securing MCP for enterprise adoption is a good refresher before the auth sessions tomorrow.
