Communication as a design choice
ProtocolBench compares A2A, ACP, ANP, and Agora under common task and runtime conditions. Protocol-normalizing adapters, controlled non-protocol factors, and unified logging make their trade-offs comparable. The four evaluation axes are task success, end-to-end latency, message or byte overhead, and robustness under failure.
The scenario suite includes GAIA, Streaming Queue, Fail-Storm Recovery, and Safety Tech. These settings expose different demands: completing a task, sustaining a stream of work, recovering from disruption, and handling safety-sensitive interactions.
Adaptive protocol selection
ProtocolRouter learns to select a protocol per scenario or module from requirements and runtime signals. ProtocolRouterBench supplies a standardized routing evaluation. The study treats protocol choice as a context-dependent systems decision, not a universal ranking.
These are communication protocols. Tool-invocation standards and RELIC’s learned organizational rules address different layers; they should not be treated as the same mechanism.
Paper and contribution
Hongyi Du is the first author. This brief follows ProtocolBench: Which LLM MultiAgent Protocol to Choose?, including the current ICML 2026 camera-ready record. The paper provides experiment-specific results; the repository provides benchmark artifacts.
Loading comments…