AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 26, 2026

The Agentic AI Foundation (AAIF) announced four new projects at AGNTCon + MCPCon North America 2026, held October 22-23 at the same San Jose venue that hosted the PyTorch Conference. The additions take the foundation from three founding projects to ten within its first year. Alibaba contributed OpenSandbox, a secure sandbox for running agents. OpenAI contributed Open Responses, an open specification for interoperable LLM interfaces. Anthropic added Agent Skills, and Hugging Face added OpenEnv, a framework for reinforcement learning environments used to train and evaluate agents. Executive director Mazin Gilbert opened the event with the project news and three new governance committees. CTO Manik Surtani framed the milestone around the foundation's original bet that agentic AI needed a neutral home, noting that a year ago AAIF launched with three projects.

Surtani's talk, "Who Keeps It Open?", argued that a license by itself does not keep infrastructure open. He contrasted Android, which stayed under one company, with Kubernetes, which Google moved into a neutral foundation. In his reading, the open problems for agent infrastructure now sit in identity, attestation, audit and payments. Outside coverage was more skeptical. The New Stack observed that Open Responses already has backing from AWS Bedrock, Databricks, Ollama, Nvidia and vLLM, but that Anthropic still defines its own Messages API and Google its Interactions API. The Register characterized the foundation as friendly and funded and cautioned that rapid embrace of AI agents may be too fast for some industries.

The protocol itself dominated much of the program. David Soria Parra, co-creator of MCP at Anthropic, argued that capable models now need connectivity more than help staying on task. The July 28 specification added stateless servers, which some developers worried would reduce MCP to just another API; some platforms already see roughly 70% of their traffic in that mode. Soria Parra announced MCP tasks support in Claude Code, allowing tool calls to run for minutes, hours or days rather than seconds. Anthropic also donated the Agent Skills standard to the foundation. The project roadmap names five priority areas, spanning agent messaging and skills over MCP, identity and an HTTP-native transport.

Production experience with MCP servers came from OpenAI's Thomas Sarlandie, who drew on the ChatGPT plugin directory. There is no practical limit on tool count, but oversized schemas get trimmed, beginning with argument descriptions. About three in four tool schemas on published servers are under 1 kB, while the largest reach about 260 kB for a single tool and 1.4 MB for a server. OpenAI rejects submitted servers unless every tool carries annotations, which let ChatGPT fast-track read-only calls and require confirmation for destructive ones. Only 56% of plugins use server instructions, which Sarlandie said can safely run to around 1,000 characters. In practice, these figures give server authors a usable rule of thumb: small schemas and complete annotations are not stylistic preferences but adoption requirements.

Chrome's Sarah Drasner presented WebMCP, a proposed web standard now in origin trial that lets websites expose tools to agents. Agents call declared tools instead of scraping the DOM, and Drasner cited benchmarks showing actions 2.5 to 7.5 times faster. Read-only and untrusted-content hints help agents resist prompt injection hidden in user comments. Drasner stressed that WebMCP complements MCP rather than replacing it, since it standardizes the browser-side surface while MCP covers server-side connectivity.

Security was the sharpest thread. UC Berkeley professor Dawn Song argued that frontier AI will help attackers more than defenders in the near term. Her CyberGym benchmark spans more than 1,500 vulnerabilities across nearly 200 open source projects, and models now find zero-days. On ExploitGym, GPT-5.6 Sol produced about 230 working exploits against 869 targets in two hours, and still more than 80 with standard security mitigations enabled. Because an attacker needs only one hole, Song called for secure-by-construction systems with AI-generated proofs and pointed to a new observatory tracking these capabilities.

Anthropic's Den Delimarsky, the protocol's lead maintainer, said cumulative SDK downloads have reached 4.17 billion, up from about 3 billion a month earlier, which he said raises the stakes of every security gap. About a third of the 116 conformance scenarios now test security, and enterprise-managed authorization is live in Claude. His auth roadmap marks that feature stable, places proof-of-possession tokens and workload identity federation in draft, and ends with agents that sign in as themselves rather than borrowing a user's token.

Agent identity also showed up in product form. Block's Bradley Axen, creator of goose, demonstrated Buzz, an open source workspace where people and agents share channels. Each agent gets its own identity plus an attestation naming its owner, and can read channels but not direct messages. Buzz reports 80,000 communities and 36,000 GitHub stars, with version 1.0 due in mid-November. Thomas Dohmke, former GitHub CEO and now CEO of Entire, said agents have turned code review into the bottleneck, citing industry data that teams using coding agents produce about 57 pull requests a week against roughly 10 for others. He wants agent session logs stored in Git and linked to commits so reviewers can check a change against its intent.

Tim O'Reilly called MCP an "architecture of participation" for AI, in the tradition of Unix, and argued the AI economy still lacks its mechanisms the way the web needed pay-per-click advertising and secure payments. He framed the fix as mechanism design, or reverse game theory: start from the desired outcome and work backward to the rules that produce it, moving trust out of the model and into protocols, gateways and verifiable identity. J.R. Storment, executive director of the 11-week-old Tokenomics Foundation, told the audience to stop counting tokens, since cost per task beats cost per token and cost per outcome is better still, but business value is the real measure.

The practical takeaways for teams running MCP servers were specific: annotate every tool, keep schemas small, and prepare for agents carrying their own identity instead of a borrowed user token. Several figures were stage claims without published data, and some security extensions remain drafts. The next AGNTCon + MCPCon events are scheduled for San Francisco on April 28-29 and London on September 15-16, 2027.

Why it matters: Teams operating MCP servers now have concrete adoption constraints from OpenAI's directory plus a security roadmap shifting toward agent-specific identity, and my inference is that borrowed user tokens will increasingly look like a legacy pattern. Anthropic's move of Agent Skills into the foundation and the draft auth work suggest consolidation around shared standards, even as competing vendor APIs remain. Developers building agent products should treat schema size, annotations and identity design as near-term compatibility work rather than future cleanup.

Read at InfoQ · AI, ML & Data Engineering

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

At AGNTCon + MCPCon North America 2026, the Agentic AI Foundation added OpenSandbox, Open Responses, Agent Skills, and OpenEnv, reaching ten projects in its first year. Keynotes from Anthropic, OpenAI, Google, and UC Berkeley focused on MCP in production, security, and agents with their own identities. By Andrew Hoblitzell