What happened
The Agentic AI Foundation (AAIF) announced the winners on October 23, 2026, in a blog post by executive director Mazin Gilbert. The awards were announced at AGNTCon and MCPCon North America in San Jose. AAIF says nominations opened in June and were reviewed by a volunteer Selection Committee drawn from the Governing Board and Technical Advisory Council, with participation from the wider community. The foundation also introduced the Keystone Award as its highest technical recognition.
The Open Agentic AI Innovation Award recognizes organizations that deployed or advanced AAIF open standards in production, research, or public tooling with real-world impact. The winners are AURA from Mezmo, Opaque Systems, and Vstorm. AAIF describes AURA as Mezmo’s open-source agent harness for production site reliability work. It connects to MCP servers and lets operators give AI agents a scoped set of tools rather than expose every available tool to every agent. The announcement presents this as a way to make production systems safer to operate with AI in the loop; it does not provide deployment metrics or a detailed access-control design.
Opaque Systems built AgenTrust, described by AAIF as an open-source tool for proving what an AI agent actually did in production. The foundation says that in its first three months, AgenTrust recorded close to half a million package downloads, with teams continuing to build integrations on top of it. The announcement does not specify how downloads were counted, how many distinct teams use the tool, or what evidence it records. Vstorm, an agentic AI engineering company, became the first agentic AI engineering and services company to join AAIF in December 2025, according to the post. AAIF says Vstorm puts AGENTS.md into every project it ships and built MCP support into its own open-source framework, so project knowledge guiding coding agents during development also guides agents in production.
The other awards recognized individual and community contributions. Nate Barbettini and Marcelo Trylesinski received the Outstanding Technical Contribution Award for work on MCP authorization and the MCP Python SDK, respectively. The Community Builder Award went to the organizers of MCP Meetup Tokyo—Satoshi Ito, Tomotada Sonoda, and Masaki Tsukada—and Lin Sun. Den Delimarsky, an MCP lead maintainer at Anthropic, received the first Keystone Award for work spanning technical contributions, security, and community. The post names contributors and activities, but gives no award scoring, financial prize, or technical evaluation criteria.
What the archive adds
New Runtime’s earlier coverage on October 22, 2026, described 1Password’s Unified Access Pro as a platform for discovering, securing, and auditing access for humans, AI agents, and machine identities. This award announcement is a new event, but its attention to scoped tools and agent activity records touches the same operational concern: how organizations govern agent access and actions.
Trend signal
This supports the broader direction described in Agent security moves to runtime boundaries: AURA’s scoped tool access is a runtime control, while AgenTrust focuses on evidence of actions after agents operate. The examples are suggestive rather than proof of a market-wide shift; AAIF’s announcement does not compare these approaches with prompt-based controls or measure their security outcomes.
Why it matters
The three projects describe different pieces of an operational stack, not a single standard recipe. Scoped MCP tools address what an agent can reach; an activity-evidence tool addresses what can be examined afterward; project context carried into runtime addresses whether deployed agents act with the same knowledge as development agents. Teams building agent systems can use this as a checklist of questions: what tools are exposed, what records are retained, and which project instructions follow the agent into production?
The announcement is not enough to select a product or establish that any control is effective. Still, the distinctions matter because a successful demonstration or a model’s instruction to behave safely does not itself establish an enforceable access boundary or an auditable record. The useful next step is to evaluate these mechanisms against a real workflow and explicit operational requirements.
Limits and uncertainty
AAIF’s post is an awards announcement, not an independent assessment of the projects. It does not publish selection scores, evaluation criteria beyond broad award descriptions, security testing, deployment details, or comparative outcomes. The nearly half-million AgenTrust package downloads are attributed to AAIF, but the announcement does not define the metric or establish active usage. It also does not report how many deployments use AURA’s scoped tools or what measurable risks they reduced. The examples show what the foundation chose to recognize, not verified evidence that the tools work equally well across production environments.
Reader takeaway
The award is a useful pointer to three practical design questions for agent deployments: can operators limit the tools an agent can use, can they establish what it did, and does the context that guides it survive the move from development to production? AAIF names projects addressing those concerns, but teams should treat the announcement as a starting point for evaluation, not as proof of effectiveness.
Open question
How do AURA and AgenTrust perform in independently evaluated production deployments, and what evidence can teams inspect before adopting them?