What happens when AI agents need to talk to each other?
Today, most agent systems — OpenClaw included — operate as single-agent setups. Your agent calls tools, manages skills, and interacts with APIs. But it doesn’t negotiate with other agents. It doesn’t discover what capabilities neighboring agents have. It doesn’t authenticate peer-to-peer.
Huawei thinks that’s about to change. At MWC 2026 in Barcelona, the company is open-sourcing A2A-T (Agent-to-Agent for Telecom) — a protocol designed to let AI agents find each other, verify trust, and collaborate on complex workflows.
What A2A-T Actually Is
A2A-T isn’t a chatbot framework. It’s infrastructure plumbing for multi-agent systems:
- Protocol SDK: Standardized message format for agent interactions — think HTTP but for agent-to-agent requests
- Registry Center: Authentication, addressing, and skill discovery — agents can advertise what they do and find agents that do what they need
- Orchestration Center: Low-code visual workflows with pre-built solution packages for common multi-agent patterns
The protocol was jointly developed with global telecom partners and released in beta (IG1453) at TM Forum Accelerate Week on February 6. The open-source launch at MWC on March 2 makes the SDK, registry, and orchestration tools publicly available.
Why This Matters Beyond Telecom
A2A-T was designed for telco networks — think automated fault detection, cross-domain service provisioning, and network optimization where multiple specialized agents need to coordinate.
But the patterns are universal:
Service discovery: Agent A needs data enrichment. Instead of hardcoding an API, it queries a registry for agents with that capability. This is DNS for AI agents.
Trust and authentication: When Agent A calls Agent B, both verify identity and permissions. No more “any tool can call any other tool” flat trust models.
Workflow orchestration: Complex tasks get decomposed across specialized agents with formal handoff protocols, error handling, and rollback.
These are exactly the problems that surface when you scale from one OpenClaw agent to many — or when your agent needs to interact with agents from other platforms.
How This Relates to OpenClaw
OpenClaw already has primitives for multi-agent coordination: sub-agents, session spawning, and tool delegation. But these are intra-system — your agents talking to your other agents.
A2A-T points toward inter-system coordination:
- Your OpenClaw agent discovering a specialized research agent on another platform
- Authenticated skill exchange between agents running on different hosts
- Standardized workflow handoffs that work across OpenClaw, NanoClaw, or any other agent framework
Google’s Agent-to-Agent (A2A) protocol and Anthropic’s Model Context Protocol (MCP) are tackling adjacent problems. A2A-T adds telecom-grade reliability requirements — exactly the kind of guarantees enterprises need.
The Standards Race
We’re now seeing multiple competing approaches to agent interoperability:
| Protocol | Origin | Focus |
|---|---|---|
| MCP | Anthropic | Tool/context integration for single agents |
| A2A | General agent-to-agent communication | |
| A2A-T | Huawei/TM Forum | Telecom-grade multi-agent orchestration |
These aren’t necessarily competitors — they operate at different layers. MCP handles how an agent talks to tools. A2A and A2A-T handle how agents talk to each other. The eventual stack probably uses all three.
For OpenClaw, the interesting question is which protocols to adopt and when. MCP support is already common in the ecosystem. A2A-T’s registry and orchestration patterns could be valuable as multi-agent setups mature.
What to Watch
The open-source launch happens March 2 at 14:30 CET at the Global Autonomous Network Industry Summit. Key things to track:
- SDK quality: Is it production-ready or proof-of-concept?
- Adoption signals: Do non-telecom platforms start implementing the protocol?
- Convergence with A2A: Will Google’s protocol and Huawei’s find common ground?
If agent-to-agent communication becomes standardized — even partially — it changes the calculus for every agent platform. Your OpenClaw agent stops being a solo operator and becomes a node in a network.
That’s a fundamentally different product. For a look at how multi-agent coordination works today in OpenClaw, see our multi-agent setup guide and the Kanban board for agent task management.
The A2A-T open-source launch takes place March 2, 2026 at MWC Barcelona. The protocol SDK, registry center, and orchestration tools will be publicly available.