A2A

Where A2A earns its keep is heterogeneity. If your org has a LangGraph agent in one team, a CrewAI agent in another, and a Google ADK agent somewhere in between, and none of them are ever going to share a runtime, a protocol beats an acquisition. You define Agent Cards, negotiate tasks over JSON-RPC, and each team keeps its own stack. The 36% token reduction from the lossless codec is the detail I'd watch if cost is the blocker. Inter-agent chatter multiplies fast — one user request fans out into dozens of messages — so a third off message tokens is not a rounding error at volume. For teams with audit obligations, the replayable A2A jury example is the more interesting release. Tracing which agent influenced a decision across a multi-step workflow is exactly the thing that stalls enterprise agent rollouts, and having a worked example beats inventing your own tracing convention. Pass on this if you want to install something. There is no managed A2A platform here, no visual builder, no out-of-the-box agent runtime. You get a specification, SDKs in six languages, a CLI, and now reference servers in Rust and N Lang — all of which shorten the path, none of which remove the integration work. The closest alternative is MCP, and the split is clean rather than competitive: MCP standardizes how an agent calls tools, A2A standardizes how agents call each other. Most serious multi-agent shops will end up running both. If your problem is one agent that needs more tools, you don't need A2A yet. In practice the caveat is governance churn. The move to the Agentic AI Foundation is a positive signal for vendor neutrality, but standards under active re-homing tend to move, and you should expect to track spec revisions rather than pin a version and forget it.