The Protocol War Is Over. Nobody Told Your Agents Who's in Charge.
A2A just joined MCP under Linux Foundation governance — and a gap analysis of five agent protocols shows every one of them is silent on the thing that actually kills multi-agent systems in production.
Your Agents Now Speak the Same Language. They Still Can't Agree on Anything.
Google handed A2A to the Linux Foundation this month, parking it next to Anthropic's MCP inside the Agentic AI Foundation. Every hyperscaler and lab that matters is in the room. If you've been stalling your agent roadmap waiting for the standards war to settle — congratulations, it settled, and you no longer have an excuse. At Kuaray, we've been telling clients since March that MCP was the winning bet. That call aged well. The uncomfortable part is what the consolidation reveals.
What Actually Got Standardized
Two protocols, two layers, and they don't overlap:
| Layer | Protocol | What it settles |
|---|---|---|
| Agent → tool | MCP | How a model reaches your DB, your search, your internal API |
| Agent → agent | A2A | How independent agents talk across trust boundaries |
Neutral stewardship, a real TSC, no single vendor's roadmap dictating the spec. That's genuinely good news, and it means the integration tax on your agent stack just dropped permanently.
Now the bad news.
The Gap Nobody Put in the Press Release
A recent arXiv gap analysis (2606.31498) ran MCP, A2A, ACP, ANP, and ERC-8004 against a six-dimension governance taxonomy: membership, deliberation, voting, dissent preservation, human escalation, and audit/replay. The verdict is brutal. Voting and dissent preservation are universally absent across all five. Deliberation is absent or partial everywhere.
Read that again. You can now wire twelve agents together with zero integration friction, and there is no standard way for them to disagree, escalate, or leave a record of who overruled whom.
This is the failure mode we keep getting called in to clean up. The pilot works. The demo dazzles. Then someone asks "which agent decided to issue that refund, and on whose authority?" — and the honest answer is a CloudWatch log and a shrug. AWS Bedrock AgentCore ships a registry, Cedar authorization, CloudTrail audit, and still encodes no trust scoring, no behavioral reputation, no capability verification. It's not a product gap. It's a layer that nobody has built yet.
To be fair: none of these protocols ever claimed to be a governance layer. That's the point. The industry standardized the plumbing and left the accountability to you.
What This Means Monday Morning
Regulators noticed before most CTOs did — NIST is scoping agent identity and permissions, Google shipped AP2 for agent payments, and the AI AGENT Act (S.5051) is live. The direction of travel is verifiable, task-bounded authorization records. Build toward that now, not after the audit.
Three things we'd do this quarter:
- Adopt MCP and A2A without hesitation. The bet is de-risked. Migrate off bespoke connectors.
- Write your governance layer yourself. Per-agent identity, scoped and expiring capability grants, an immutable decision log, and a mandatory human-escalation path for anything irreversible. Nobody is shipping this for you.
- Treat multi-agent consensus as a design problem, not an emergent property. If your architecture has no explicit mechanism for dissent, your agents will resolve conflicts by whoever writes last. That's not consensus. That's a race condition with a budget.
Schedule a Technical Architecture Review with our Strategists — we design the governance layer the protocols left out.