← /editions
#70

A2A found a home. The stack is complete.

// Agent2Agent joins the Agentic AI Foundation as a hosted project alongside MCP, completing the open agentic stack under neutral governance — the same fortnight Check Point disclosed 11 vulnerabilities across six major agent frameworks.

$ today_in_10_seconds

A2A — the protocol that lets agents built by different vendors discover each other and hand off work — is joining the Agentic AI Foundation as a hosted project, putting it under the same neutral governance as MCP. The open agentic stack now has every layer covered: instructions, runtime, tools, traffic, and agent-to-agent. The uncomfortable part is that the frameworks actually implementing those layers just took 11 disclosed vulnerabilities.

$ lead_story

A2A joins AAIF's open agentic stack

Agent2Agent — launched by Google in April 2025, donated to the Linux Foundation, and merged with IBM's Agent Communication Protocol in August 2025 — is now a hosted AAIF project alongside MCP, goose, agentgateway, and AGENTS.md. The division of labour is clean: MCP standardizes how agents reach tools and data; A2A standardizes how independent agents discover one another, delegate, and exchange results across organizational boundaries. This isn't a paper spec looking for adopters — v1.0 shipped in March 2026 with signed agent cards, multi-tenancy, and version negotiation, and it's running in production at Huawei (as the protocol between its OS-level assistant Celia and in-app agents on HarmonyOS), inside WeChat, and natively across Google Cloud, Azure AI Foundry, and AWS Bedrock AgentCore. Google Cloud and PayPal are extending it into commerce, with shopping and merchant agents negotiating over A2A and the Agent Payments Protocol handling authorization. For builders, the practical change is governance, not capability. A2A's 150+ backers include direct competitors, and that only holds together under a structure no single participant controls. If you were hedging on A2A because it was still functionally Google's, that objection is gone.

// why_it_matters: The cross-vendor communication layer your multi-agent architecture depends on is no longer subject to one company's roadmap.

read more →

$ also_worth_attention

  • Check Point breaks six agent frameworks — and none of it was novel

    At Black Hat, Check Point Research disclosed 11 vulnerabilities across LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework, and Google ADK after a year of testing. The bug classes were insecure deserialization, SSRF, path traversal, use-after-free — problems the industry learned to fix twenty years ago, now sitting underneath agents that read your inbox and write to your database. The insight that should reorganize your threat model: prompt-controlled content can cross the boundary into trusted framework logic itself, which means your input filtering was never the whole job. The agent needs no dangerous tools to be turned against you — reading the wrong document is enough.

    link →
  • Cloudflare ships Kitesurf, a browser that doesn't pretend agents are people

    Cloudflare's argument is that agents don't care about tabs, themes, extensions, or 60fps scrolling — they care about token count, context windows, scalability, and cost. So Kitesurf runs entirely on Workers and uses 3–7x less CPU and memory than Chromium for screenshots and HTML extraction, though it takes 1.7 to 1.8 times longer to finish those tasks. That's the actual trade: you're buying concurrency and unit economics, not latency. It's free in beta via Browser Run and speaks CDP, so Playwright and Puppeteer clients work unchanged.

    link →
  • The UAE is building sovereign agent infrastructure, not buying seats

    Federal officials reviewed the Agentic AI Project's five tracks this week, including FedAI — a shared federal platform for building agentic solutions — and a National AI Gateway giving entities single-point secure access to multiple AI platforms. Entities will be able to run commercial models alongside open-source models hosted on sovereign infrastructure, with a classification and assessment scheme for the agent systems themselves. The target is 50% of federal services and operations within two years, under a 'human leads, AI enables' framing. Whatever you think of the timeline, the architecture — gateway, model portability, classification tiers — is a better reference design than most enterprise agent programs currently have.

    link →

$ bigger_picture

The trend behind the headlines

Watch where the standardization is happening. A2A, MCP, and Agent Plugins are all consolidating the seams — how agents find each other, reach tools, and ship capability bundles. Nobody is standardizing what happens inside the box, and Check Point just showed why that gap costs you: the orchestration layer is a fresh attack surface running decade-old bug classes with production credentials attached. The seams maturing faster than the implementations is the defining condition of this moment, and it's exactly the condition under which governments and enterprises start committing to two-year rollouts. Treat your agent runtime as a privileged identity with full credential access, because that's what it is. The protocol politics are settling. The security posture isn't.

$ tool_of_the_day

Agent Plugins 1.0

An open, vendor-neutral packaging format published August 6 that bundles Agent Skills and MCP server configs into one portable directory: a plugin.json manifest, an optional skills/ folder, and an optional mcp.json. Path-containment rules are normative and credentials can't be embedded, so package validation is part of interoperability rather than a per-client guess. Migration is close to free — add a manifest, move existing skills, declare your servers.

explore Agent Plugins 1.0 →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime