Description

MMA2A extends A2A with modality-native routing: instead of converting all inter-agent communication to text, the layer inspects Agent Card capability declarations and routes voice, image, and text parts in their native formats.

Results

• 52% task completion vs 32% text-only baseline (+20 pp) on CrossModal-CS benchmark
• Vision-dependent tasks: +38.5 pp (defect reports), +16.7 pp (visual troubleshooting)
• Trade-off: 1.8× latency from native multimodal processing
• Key finding: routing alone is insufficient — downstream agents need capable reasoning to realize the gains

Relevance to Zeph

Zeph's zeph-a2a currently serializes all inter-agent messages as text/JSON. MMA2A demonstrates that adding capability-aware routing to AgentCard declarations enables non-text modalities (image, audio) to pass natively between agents.

Immediate actionable improvements:

1. Add capabilities field to Zeph's AgentCard / agent discovery (already partially supported)
2. When routing a task, inspect calling agent's capability declarations and preserve binary payloads
3. No need for full MMA2A — even step 1 (declaring image/audio capability) enables future routing

Architecture Sketch

Source

• Paper: https://arxiv.org/abs/2604.12213
• Benchmark: CrossModal-CS