AgentMail
Day 34 after signup
The State of A2A: What Died and What Survived - AgentNews #36
IBM's ACP lasted five months. Google's A2A is still around, barely. Here is what actually won.
Why it works
- Open with a timely industry question, then use concrete examples to make the broader analysis easy to follow.
- Move from the problem to a practical alternative before introducing the product, so the pitch follows naturally from the argument.
- Keep the main call to action focused and close with a concise feature spotlight for readers who want a product detail.
How it is built
- Issue masthead
- Protocol landscape
- Adoption analysis
- Email as alternative
Email copy
New: polling webhooks. Inbox events with no public URL → Issue 036 · November 2026 The State of A2A: What Died and What Survived The other day, I stumbled onto a tweet announcing IBM's Agent Communication Protocol. Remember that one? IBM shipped ACP in 2025. I was pleasantly surprised by this. It was probably IBM's attempt to get hip with the new kids in town. A month later, Google announced its own agent protocol, A2A, with 50+ partners lined up behind it. Where are all these protocols now? Nowhere, really. ACP lasted about five months before IBM folded it into A2A and said the team would be winding down development. A2A is still active, kind of, but no one's actually using it. When the Linux Foundation marked A2A's first birthday, they counted supporters and GitHub stars, but not how many agents were actually talking over the thing. So, what happened? For the most part, it's back to agents and MCP now. My read is that these protocols were more futuristic than what we actually have, which is probably what stalled adoption. They were too complicated and divorced from the actual use cases at the time, and even now. They were written for swarms of agents and 50-agent orchestrations, a world where agents could look up any other agent and just start talking and negotiating. That world was never really on the books at that point, and it still seems a bit distant. What was actually happening at the time was one person running a coding agent, or a personal assistant, or a customer support agent. All of these use cases needed to reach people and services that already existed. A2A did not do discovery well. One of the fundamental flaws was that most of these protocols only helped you reach an agent whose domain you already knew. Then there was Moltbook. Moltbook was a social network built only for agents. Humans could watch but not post. It went viral and grew fast right away, and Meta bought it in March. In our opinion, this may have actually slowed down agent adoption. This was the purest version of the dream: a place built from scratch for agents to talk to agents. The thing is, it ended up being a place where people puppeteered their agents to do what they wanted. After all, we did it ourselves. All of these protocols and platforms made the same mistake. They built somewhere new and asked agents to move in. Think of it this way: when the world was first introduced to cars, we didn't build new roads for them. They drove on the same roads that horse-drawn carriages used, and we improved the roads later on. The point is, we needed to onboard agents to the places and services we already work in. What survived? MCP is the clear winner. It never asked anyone to join a new network. It put tools inside Claude, Cursor and coding agents, places people already had open all day. MCP let you connect to many things in minutes, while with A2A you had to spend days learning orchestration concepts. But there's an even simpler way, which is email. It sounds biased, but look at the last two months. All these personal agent providers started building email and introduced email services for their agents. Muse Mail and Grok Bot's mail are cases in point. Even Google's Gemini coworker agents got addresses inside Workspace. Google, the company that originally wrote A2A, didn't even mention it in the launch. When Google needed its agents to be reachable, it gave them an inbox. Email already has what the new protocols were trying to invent. Over 4.7 billion people already have an address, and trust is anchored to domains. Once two agents have addresses, they can talk without anyone writing a spec. We've watched two OpenClaw agents build an HTML5 game together on a single email thread. This was one of the earliest signs that email was going to be the primary, and the easiest, way for agents to talk to each other. Don't rent your agent's identity, though This is the dangerous trend, in my opinion. A mail.grokbot.com address belongs to Grok B