AGNTCon + MCPCon 2026: Agentic AI Foundation Adds Four Projects as MCP Turns to Security

The Agentic AI Foundation expanded its scope at AGNTCon + MCPCon North America 2026, growing from three initial projects to ten in its inaugural year by adding OpenSandbox, Open Responses, Agent Skills, and OpenEnv.

The event, held October 22-23 in San Jose, shifted its focus from new protocol features toward security and identity, with keynotes from Anthropic, OpenAI, Google, and UC Berkeley addressing MCP in production.

Mazin Gilbert, the executive director of the AAIF, highlighted the project additions and three new governance committees. Alibaba contributed OpenSandbox, a secure sandbox for running agents, while OpenAI provided Open Responses, an open specification for interoperable LLM interfaces.

Manik Surtani, the foundation's CTO, remarked on the milestone, stating, "A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10." He stressed that a license alone is insufficient to maintain open infrastructure, contrasting Android with Kubernetes.

David Soria Parra, co-creator of MCP at Anthropic, contended that advanced models now require connectivity more than assistance with task completion. He announced MCP tasks support in Claude Code, allowing tool calls to run for extended periods, unlike previous seconds-long limits.

OpenAI’s Thomas Sarlandie detailed what the ChatGPT plugin directory reveals about MCP servers in production. He noted that while there is no practical limit on tool count, oversized schemas are trimmed, and OpenAI requires every tool to carry annotations for fast-tracking read-only calls.

Chrome's Sarah Drasner introduced WebMCP, a proposed web standard currently in origin trial, enabling websites to expose tools to agents. Benchmarks cited by Drasner showed actions executing 2.5 to 7.5 times faster when using this standard.

Dawn Song, a UC Berkeley professor, cautioned that frontier AI may initially aid attackers more than defenders. Her CyberGym benchmark demonstrated that models can now find zero-days across nearly 200 open-source projects.

Anthropic's Den Delimarsky reported that cumulative SDK downloads have reached 4.17 billion, noting that about a third of the 116 conformance scenarios now test security.

Thomas Dohmke, former GitHub CEO, observed that agents have made code review a bottleneck, citing industry data showing teams using coding agents produce about 57 pull requests weekly, compared to roughly 10 for others.

Tim O'Reilly characterized MCP as an "architecture of participation" for AI, suggesting the necessary mechanisms for the AI economy must be designed, moving trust into verifiable identity and protocols.