The open agentic AI stack is being built now. AAIF Ambassador Marco Gonzalez breaks down four key projects shaping the ecosystem—from protocols and runtimes to conventions and infrastructure. MCP. Goose. AgentGateway. AGENTS.md. Learn more: https://bit.ly/4uXVDab

Countdown to recognizing the builders of open agentic AI. The Agentic AI Foundation AAIF Annual Technical Excellence Awards highlight contributors across MCP, goose, AGENTS.md, and agentgateway. Nominations open until July 17. Submit: https://bit.ly/3QtI04h

Good ideas come from people who show up. We are launching IRL meetups for the Agentic AI Foundation to bring together people building and shaping agentic AI in real time. Just builders sharing what they are learning as they go. Apply here: https://bit.ly/4eHIVGj

Meet goose. goose shows what agent systems look like when they move beyond chat. As an Agentic AI Foundation project, it’s built on open standards like MCP-based extensions, connecting LLMs to tools, execution, and recovery loops. https://bit.ly/4oLtodo

Communities are what turn open source projects into movements. The AAIF Ambassador Program brings together 139 ambassadors across 41 countries who are helping grow the open agentic AI community through education, collaboration, and hands-on contributions. Meet the ambassadors: https://bit.ly/4g6nqBv

Agent discovery is emerging as a foundational layer of the agentic ecosystem. Google’s Agent Resource Discovery (ARD) introduces an open specification for publishing, discovering, and verifying agentic resources across tools, APIs, MCP servers, skills, and agents. https://bit.ly/4uVzqts

What is the right deployment unit for an AI agent? AAIF Ambassador Lin Sun explores how agent systems may need new approaches to deployment and architecture as the agentic AI ecosystem evolves. Read on: https://bit.ly/4vzjKx4

The goose team uses Terminal-bench and Harbor to surface failures, but humans are responsible for interpreting results. Douwe Osinga, goose maintainer, notes that benchmarks are most useful when treated like bug reports rather than trusted scoreboards. Read on: https://bit.ly/4vq2vy5

The Agentic AI Foundation has launched the AAIF Annual Technical Excellence Awards to recognize contributors building open agentic AI infrastructure, including work across MCP, goose, AGENTS.md, and agentgateway. Submit nominations by July 17: https://bit.ly/3QtI04h

Kilo proposed REVIEWS.md, a repo-level standard for agent-native code review. It lets review agents understand project-specific coding standards, architecture choices, and team norms. Better context leads to better code reviews: https://blog.kilo.ai/p/code-reviews-md

In this Community Spotlight, Katherine Druckman, Head of Community and Partnership Engagement at JetBrains, shares why open collaboration, interoperability, and community-driven standards are essential for moving AI agents into production. Read more: https://bit.ly/4xP0f52

38% to 87% tool selection accuracy. Samyuktha Mohan Alagiri from IBM explores multilingual MCP and the challenges of tool calling across Indic languages. The session covers language-aware routing, dual-language tool manifests, and localization patterns: https://bit.ly/4uW50XX

Solo.io published a reference architecture using agentgateway as an in-path control layer for LLM and agent traffic. It enables real-time cost controls across agentic AI workloads at the gateway level: https://www.solo.io/blog/building-real-time-ai-cost-controls-with-agentgateway

Jake Pineda (Head of Membership Development, AAIF) keeps hearing a consistent theme from members: strong technical work is already happening, along with clear ideas on where agentic AI standards should go next. AAIF is building on that momentum by making engagement clearer: https://bit.ly/4eCcFV7

We spent decades designing software for humans. Now machines are becoming the users. Sam Partee from Arcade explains the shift from UX to MX (Machine Experience) and why MCP developers need to rethink how tools are designed, described, and discovered by AI systems. Watch: https://bit.ly/44gw0Gz

Stripe shared lessons from steering AI agents: Agents often miss subtle hints, buried warnings, and adjacent documentation. Direct prompts, explicit errors, and workflow-level instructions work better than relying on context. https://bit.ly/3QqQrxp

Today, we're welcoming 138 ambassadors — spanning 41 countries and 6 global regions. What they've committed to: One public contribution per month, tied directly to an AAIF project, useful to a real developer. Read more from AAIF's VP of Developer Experience, Angie Jones: https://bit.ly/4uP7mHZ

Agentgateway 1.3.0 ships with a strong focus on production LLM operations: an LLM-first UI, token + dollar cost tracking, and virtual models for weighted, conditional, and failover routing. Read on: https://bit.ly/4enpe82

Phillip Lamb (ClickHereLabs, Live Oak Digital) and Liad Yosef (Ora, MCP Apps) show MCP and agentic web standards are converging on structured tool surfaces as the real interface layer. Accessibility and agent-readiness are becoming the same engineering problem. Read on: https://bit.ly/4b2mZok

MCP works on your laptop. But what happens at internet scale? Prashanth Subrahmanyam from Google Cloud breaks down the “production cliff” — from unused MCP connections and stateful bottlenecks to the protocol changes helping MCP scale. For anyone building agents: https://bit.ly/4uDwNwh

AWS open sourced Agent-EvalKit, an Apache 2.0 toolkit for evaluating AI agents. Instead of only judging final answers, it examines the full workflow: tool use, execution traces, faithfulness, and response quality. Read how production agents need evaluation beyond the output: https://go.aws/4vIfkDM

MCP development is shifting from tool wiring to full application design. Skybridge, launched on Product Hunt, introduces a full-stack TypeScript framework for building MCP and ChatGPT apps where tools, UI components, and agent workflows are defined together in a single system. https://bit.ly/4emNYxc

Google Cloud introduced Open Knowledge Format, an open spec for portable agent context. It represents organizational knowledge as Markdown with YAML frontmatter so it can be shared, versioned, and consumed by different agents. Read on: https://bit.ly/4el1srR

The latest episode of The Context featured Chris McHenry, CPO at Aviatrix, discussing: Aviatrix + Obot: The Containment Era Launch Hear how the MCP ecosystem is evolving and what builders are creating with the Model Context Protocol: https://bit.ly/4vaMpbn

AI quality issues in agentic systems are often process issues amplified by AI. Angie Jones describes shifting code review earlier in the workflow and moving guardrails into system design instead of individual developer memory. Full episode: https://apple.co/44hSbMB

Databricks open sourced Omnigent, an Apache 2.0 licensed “meta-harness” for multi-agent development. It helps teams compose agents into workflows, manage permissions and budgets, and collaborate across shared agent sessions. Read on: https://bit.ly/4ozXUqI

What does enterprise-ready MCP look like? Stacklok shares guidance on building a local MCP registry and securing MCP deployments for Kubernetes platform teams. As adoption grows, governance is becoming a key part of the MCP conversation. Read more: https://bit.ly/4oprlvh

MCP servers can expose tools. The next challenge is exposing the knowledge needed to use them well. Angie Jones explores Skills Over MCP and how Skills could provide agents with operational guidance. The goal is better discovery and more reliable agent workflows: https://bit.ly/4oBaGoP

AI agents fail in production when they lack shared project context. Antoni Kozelski, CEO and Founder of Vstorm, embeds AGENTS.md generation into every FastAPI and Next.js scaffold. 830K+ downloads later, AGENTS.md is becoming a shared standard across the Python ecosystem: https://bit.ly/4guffPz

Nick Cooper (Member of Technical Staff at @OpenAI and MCP Steering Committee) on MCP: It’s not just tool use. It’s about making agent systems secure, observable, and production-ready. Without shared foundations, every team rebuilds auth and integration patterns from scratch: https://bit.ly/4vLGdqo

What happens when MCP servers can work with other MCP servers? Jerome Swannack, one of MCP’s original architects, explores “second-order” MCP servers. Client-side routing, centralized permissions, reusable tools — the next evolution of MCP is taking shape: https://bit.ly/4eupTTL

What failure modes emerge when agents interact with tools, systems, and other agents? Microsoft updated its agentic AI failure-mode taxonomy after a year of red teaming, adding categories like goal hijacking, supply-chain compromise, and confused deputy attacks. Read on: https://bit.ly/4v56cZZ

Who is building the future of agentic AI? The AAIF Technical Excellence Awards recognize the contributors, community builders, and organizations advancing MCP, Goose, AGENTS.md, and agentgateway. Nominations close July 17! Learn more here: https://bit.ly/4fKiGkV

Today, the Agentic AI Foundation is launching the inaugural AAIF Annual Technical Excellence Awards — celebrating the individuals, teams, and organizations advancing open standards and open source projects including MCP, Goose, AGENTS.md, and agentgateway. Learn how: https://bit.ly/4etSdWD

Context windows are not memory. When a session ends or context fills up, most agents lose the history of what they've learned. Angie Jones states that this makes long running tasks and ongoing collaboration difficult. Persistent, retrievable memory is a core requirement for useful AI agents.

57 sessions. One growing MCP community. MCP Dev Summit Bengaluru brought together builders sharing what they’ve learned from building agentic systems in the real world. From tool design and security to agent-to-agent trust, here are the lessons shaping the next phase of MCP: https://bit.ly/3StNjkS

A2A tagged v1.0.1 on May 28. The release adds Rust as the sixth official SDK alongside Python, JS, Java, Go, and .NET, ships a multi-tenancy implementation guide clarifying tenant-field semantics, and aligns push-notification auth examples with actual protocol behavior: https://bit.ly/4vqH6EI

Tool use is still one of the hardest problems in agent systems. Arcade Labs launches with Omni, an open source tool router that helps agents discover tools, generate calls, and execute actions across connected services. Read on: https://bit.ly/4gdxusl What should agent-tool interfaces look like?

Cloud-based agents are not an option for every environment. Hoseong Lee from Korea Aerospace Industry Research Institute explores why goose is a practical choice for air-gapped industrial AI deployments: https://bit.ly/4vfraFC

AI agents need more than model access. They need governance, interoperability, and enterprise integration. Goose v1.37.0 adds hooks for behavior control, enterprise AI gateways, MCP integrations, and ACP interoperability: https://bit.ly/4oDuQ1z What's still missing from the enterprise agent stack?

Tighter security for every MCP session. MCP Python SDK 1.27.2 binds sessions to the authenticated principal and scopes experimental tasks to their session — closing leakage paths. AccessToken adds subject and claims fields for cleaner authorization: https://bit.ly/4fV5tFP

What if you never got woken up at 2 AM for an alert again? Diamond Bishop (Director of Eng/AI, Datadog) on agents in production: Bits AI SRE investigates alerts autonomously, Bits AI Dev fixes code from live errors, Security Analyst (now GA) handles security investigations. https://bit.ly/4dYEC9s

What causes more MCP incidents: malicious actors or missing guardrails? Cecilia Liu of Docker outlines a three-layer approach to MCP governance: gateways, sandboxed runtimes, and integration with enterprise identity and observability systems: https://bit.ly/44b42vP

Most docs are written for humans. What about agents? Member company Microsoft's Learn MCP Server gives any MCP-compatible agent live access to learn docs across terminals, IDEs, and custom assistants — a model for making docs and SDKs agent-ready. Read more: https://bit.ly/4x2WOYe

What does it take for coding agents to stay useful over longer runs? Cursor Ships Composer 2.5, a stronger model tuned for sustained coding work, with signals that an even larger model is already in development. From Cursor on X: https://bit.ly/4vhO9zo

Member company Google open sourced a framework that brings AI agents into Chrome Enterprise security operations — letting teams investigate threats, automate analysis, and query browser telemetry through agent workflows. Read more: https://bit.ly/4o2jdRm

What does “memory” mean for AI agents beyond a chat history? TencentDB Agent Memory introduces a local, MIT-licensed system that combines symbolic short-term memory with a four-tier long-term structure for conversations and personas. Explore the details from MarkTechPost: https://bit.ly/4u7f8wK

Free-token era in AI may be winding down. goose maintainer Douwe argues agent systems will need to route dynamically across closed, open, and local models, with workflows, permissions, and memory managed at the agent layer to stay efficient. Explore the post by Douwe: https://bit.ly/4u6hJH9

Why do some AI agents succeed in production while others struggle with reliability and cost? Misha Lanin, Head of Growth at Port of Context, examines how different execution approaches affect agent performance in real-world workloads. Read on: https://bit.ly/3QAoEKS

How do you safely run an AI agent in production? Anthropic published a detailed look at how it sandboxes Claude — the isolation mechanisms, security controls, and operational safeguards behind running it at scale. Practical guidance for anyone deploying agents in production: https://bit.ly/4dVtR7D