Benchmarks & Proof

Claims about concurrency and resilience are cheap; these are the numbers Jido actually hits on real hardware.

🖥️
10,000
agents on 2-core, 4GB VM
⚡
< 1ms
median message latency
💾
~20KB
memory per idle agent

Single-Node Benchmarks

Agents Memory CPU
1,000 40MB 5%
5,000 180MB 12%
10,000 350MB 22%

Test scenario: Agent behavior is a simple state machine with periodic work. Measurement via :observer and telemetry aggregation over 10-minute sustained load.

Multi-Node Scenarios

< 2s
Failover time when node dies
33%
Throughput impact during outage (1 of 3 nodes)
< 5s
Agent redistribution time

Failure Behavior Experiments

Random agent crashes

Crash 10% of agents randomly per second

Result: Supervisor restarts isolated to crashed agents

Impact: No cascade failures, 99.9% uptime for healthy agents

Thundering herd

5,000 agents all request external API simultaneously

Result: Back-pressure via mailbox monitoring

Impact: Graceful degradation, no OOM

Reproduce the Benchmarks

bash 
git clone https://github.com/agentjido/benchmarks
          cd benchmarks
          mix deps.get
          mix run bench/single_node.exs

We expect developers to rerun and verify these numbers on their own hardware.

See it in Action

▶ 10,000 agents with Observer ▶ Node failover in real-time 📁 Benchmark repo