Top AI Product

If you’ve been building anything with AI agents recently, you’ve probably felt the pain. MCP gave us a solid foundation for tool discovery, but the moment you try to ship an agent to production — with real users, real money on the line, real things that can go wrong — you realize there’s a whole class of problems MCP was never meant to solve. That’s exactly the gap [Agent Action Protocol (AAP)](https://github.com/agentactionprotocol/aap/) is trying to fill.

AAP popped up on [Show HN](https://news.ycombinator.com/item?id=47235632) in early March and immediately struck a nerve. The title said it all: “MCP got us started, but is insufficient.” Bold claim, but the spec backs it up. AAP is a declarative protocol that defines the full lifecycle of an agent action — discovery, authorization, execution, observation, and recovery. Each action is a YAML manifest describing what a capability does, what inputs it needs, what it produces, what resources it touches, and crucially, how much it can cost. That last part alone makes it worth paying attention to.

What I find most compelling is the permission model. Actions declare what they need, and runtimes assign roles that grant or deny access, with everything denied by default. There are built-in hooks for human review gates, auth injection, and cost guards. If you’ve ever watched an agent accidentally burn through API credits or modify something it shouldn’t have, you know why this matters.

The spec is also transport-agnostic, which is a smart move given the current [protocol wars](https://bestofshowhn.com/) between MCP, A2A, and ACP. AAP doesn’t compete with MCP — it sits above it. MCP, REST, CLI, function calling — all valid backends. AAP defines the “what” while transport protocols handle the “how.”

It’s still early days, and the [GitHub repo](https://github.com/agentactionprotocol/aap/) is moving fast. But the fact that it tackles real production headaches — rollback behavior, checkpointing, observability — instead of just adding another way to call tools makes it worth watching closely. If you’re building agents that do more than answer questions, give the spec a read.