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

At the AGNTCon + MCPCon 2026 conference in San Jose, the Agentic AI Foundation added four new projects, bringing its total to ten. The event focused on security and identity, following the PyTorch Conference. Key additions included Alibaba's OpenSandbox for secure agent execution, OpenAI's Open Responses for interoperable LLM interfaces, Anthropic's Agent Skills, and Hugging Face's OpenEnv for reinforcement learning environments.

Manik Surtani, the foundation's CTO, emphasized that a license alone cannot ensure openness, contrasting Android's single-company model with Kubernetes' neutral foundation. The foundation now grapples with identity, attestation, audit, and payments. Outside coverage noted that Open Responses already has support from major companies, but Anthropic still maintains its own API, while Google uses its Interactions API.

J.R. Storment warned against rapid adoption of AI agents, suggesting businesses should focus on cost per outcome. David Soria Parra highlighted the need for connectivity in capable models and announced MCP tasks support in Claude Code. Anthropic donated the Agent Skills standard to the foundation, with a roadmap focusing on agent messaging, skills, identity, and HTTP-native transport.

OpenAI's Thomas Sarlandie shared insights on MCP servers in production, noting no practical limit on tool count but trimming oversized schemas. WebMCP, a proposed web standard, allows websites to expose tools to agents, potentially improving performance. Security expert Dawn Song warned of the near-term threat of frontier AI to defenders, presenting CyberGym and ExploitGym benchmarks showcasing AI's ability to find vulnerabilities and exploits.

Anthropic's Den Delimarsky noted that cumulative SDK downloads have reached 4.17 billion, underscoring the importance of security measures.