Sub-Millisecond
Agent Communication
AISP runs over a WireGuard-encrypted NATS mesh. Agent-to-agent latency within the mesh averages 0.4ms — 30x faster than REST/HTTP and faster than most gRPC deployments.
Latency Simulation
Simulates a round-trip agent message through the AISP session layer on local mesh.
Protocol Latency Comparison
Average round-trip latency, p50, local network conditions
Memory Sync Latency
Time for a write to propagate to all mesh nodes (p99)
Methodology
Benchmarks run on 5-node mesh: Apple Silicon M-series (master, nodo) + Contabo VPS (AMD EPYC) + Kali + Flow. Connected via WireGuard over Tailscale.
Latency = time from message publish on sender to delivery confirmed on receiver. 10,000 samples per run, p50 reported. Top 1% outliers excluded.
Messages of 1KB payload, published at maximum rate from a single producer. Measured until first message drop or 30 seconds elapsed.
Write committed on one node, time measured until read confirmed on all 4 other nodes. LWW-Map with 10 keys, 100-byte values.