- Published on
AGNTCon + MCPCon 2026: Agentic AI Foundation Adds Four Projects as MCP Turns to Security
- Authors
- Name
- DeepHorus
SAN JOSE — At AGNTCon + MCPCon North America 2026, the Agentic AI Foundation added four projects, expanding from three founding efforts to ten in its first year. The October 22–23 event followed PyTorch Conference at the same venue. Much of its program shifted from new protocol features to security and identity.
The foundation launched a year ago with three projects and a bet that agentic AI needed a neutral home. Manik Surtani, AAIF’s CTO, said the milestone shows that bet is paying off. In “Who Keeps It Open?,” he argued a license alone does not keep infrastructure open, contrasting Android with Kubernetes.
Alibaba contributed OpenSandbox, a secure sandbox for running agents; OpenAI contributed Open Responses, an open specification for interoperable LLM interfaces; Anthropic added Agent Skills; Hugging Face added OpenEnv for reinforcement-learning environments. Surtani said the foundation now hosts 10 projects. The open problems sit in identity, attestation, audit, and payments.
Outside coverage was cooler. The New Stack noted support from AWS Bedrock, Databricks, Ollama, Nvidia, and vLLM, but Anthropic still defines its own Messages API and Google its Interactions API. The Register called it “vendor-affiliated AI agent tech” adopted by a “friendly, funded foundation,” warning that agent enthusiasm may be too rapid for some industries.
David Soria Parra, MCP’s co-creator at Anthropic, argued capable models now need connectivity more than task help. The July 28 specification added stateless servers, a mode some platforms already see about 70% of traffic in. He announced MCP tasks support in Claude Code, letting tool calls run for minutes, hours, or days instead of seconds.
OpenAI’s Thomas Sarlandie shared lessons from ChatGPT’s plugin directory. There is no practical tool-count limit, but oversized schemas get trimmed, starting with argument descriptions. About three in four schemas are under 1 kB; the largest reach 260 kB for one tool and 1.4 MB per server. OpenAI rejects servers unless every tool carries annotations.
Chrome’s Sarah Drasner presented WebMCP, an origin-trial standard letting sites expose tools to agents and making actions 2.5 to 7.5 times faster than DOM scraping. It complements MCP, she stressed. UC Berkeley’s Dawn Song argued frontier AI will help attackers more than defenders near term. Her CyberGym benchmark spans 1,500 vulnerabilities, and on ExploitGym GPT-5.6 Sol produced about 230 working exploits against 869 targets in two hours.
