Price-invariant
The 2-vs-29 node ratio holds at any $/node-hour. Only the absolute figure changes.
Performance / Measured on bare metal
Proof, not promises.
Same host, same workload, orchestration only. Every figure here is reproducible on commodity hardware — and the test conditions are stated so you can check them.
16×
Higher throughput
72,089 vs 4,524 /s
110×
Lower p50 latency
6.6 µs vs 726 µs
4×
Lower memory
700 MB vs 2,633 MB
12×
Faster time to ready
1.82 s vs 22.19 s
Explore the numbers
One workload. Two architectures.
Choose the metric that matters to your deployment. Lower-is-better metrics show the baseline at full width and the EAE share beside it.
Independent test harness
16×
higher throughput
One process, eight threads, shared memory. The baseline needs eight worker processes and IPC to do the same work.
Dintellinno EAE72,089 /s
Leading production framework4,524 /s
8-thread bare metal · orchestration only · LLM inference excluded
Full comparison
The whole table.
EAE versus LangGraph TypeScript, selected as the highest-performance variant of the market-leading production framework. Ratios transfer per vCPU.
Identical hardware
| Metric | EAE · Go, 8 threads | LangGraph TS · 8 workers | Advantage |
|---|---|---|---|
| Throughput verifications / second | 72,089 /s | 4,524 /s | up to 16× higher |
| Time to ready | 1.82 s | 22.19 s | up to 12× lower |
| Latency p50 / p99 | 6.6 µs / 19.5 µs | 726 µs / 2.08 ms | up to 110× lower |
| Equal-state memory per agent | 88–99 B | 128 B | ~1.3–1.5× lower |
| RSS @ 100k agents | ~700 MB | 2,633 MB | up to 4× lower |
| A2A request / response | ~2.5 µs in-process | ~41 µs IPC | ~16× faster |
| Deployment | 1 proc / 8 threads shared memory | 8 procs no shared memory | Qualitative |
Infrastructure cost
2 nodes instead of 29.
The saving is not a discount on the same architecture. It is the architecture: fewer processes, shared arenas and no per-agent heap.
Up to 94% lower compute
Model the node bill.
The 2-vs-29 ratio is price-invariant. Move the node price and see the absolute monthly difference change — the percentage stays the same.
Monthly node cost uses 730 hours, HA ×2, and 2 vs 29 nodes. Shared platform overhead is excluded. Runtime and orchestration only; LLM inference excluded.
Dintellinno EAE · 2 nodes$1,168 / mo
29-node baseline$16,936 / mo
Compute-node cost saved93%
Annual difference$189,216 / yr
Density
100,000 resident agents in one process at ~700 MB RSS, with no per-agent heap to collect.
Compute
$0.0012 per million cycles against $0.0197 for the interpreted baseline.
Methodology
How we measured.
Bare-metal Linux, 8 threads (4C × 2 SMT), 23 GB RAM. Go 1.27.1 against Node 22.22.1. Core orchestration, scheduling and inter-agent dispatch; external LLM inference is excluded from both sides.
Baseline selection
Tested against LangGraph TypeScript, specifically selected as the highest-performance, lowest-latency runtime variant of the market-leading production framework — significantly faster than Python equivalents.
What is measured
Core orchestration, concurrency scheduling and inter-agent dispatch. External LLM network inference is excluded. Ratios are transferable per vCPU.
Deployment shape
EAE runs as one process with eight threads over shared memory. The baseline runs as eight processes with no shared memory, communicating over IPC.
What this does not claim
It does not claim faster model inference, faster network calls or a faster storage layer. It isolates the runtime cost between an agent decision and the next dispatch.
Reproducing the test
We run the comparison on hardware we control with a fixed harness and publish the conditions. A pilot can reproduce the same test on your workload and your hardware, under NDA if required.
Beyond the server
Density matters at the edge too.
Native execution keeps density high and energy per workload low on limited-CPU, low-power hardware. The Rust engine carries the same contract onto ARM64 devices.
Engineering stackServer engineGo · x86-64
Edge engineRust · ARM64
Artefact<50 MB
External runtimeNone
Run it yourself
Reproduce the benchmark on your workload.
Bring a representative workload and a target budget. We will run the comparison inside your environment, not on a slide.
Orchestration only · inference excluded · conditions published