Performance Report

Status: Snapshot Last Updated: 2026-01-03 Scope: Local benchmarks for Core, Host, and World

Summary

Core + Host runs 3.5 ms to 8.6 ms avg latency per intent under compute-heavy flows.
World adds governance overhead and increases tail latency, but remains within ~9 to 21 ms p99 in tested scenarios.
Memory growth is modest under snapshot retention, and mostly attributable to retained snapshots and in-memory world history.

Environment

Node: v22.11.0
Platform: darwin arm64
CPU: unknown (os.cpus() returned empty)
Memory: 16 GB

Methodology

Benchmark script: scripts/bench-core-host-world.mjs
Deterministic context: now = 0, randomSeed = "bench"
Effect handlers: no-op (return empty patches)
Warmup before measurement to reduce JIT noise
Memory sampling uses process.memoryUsage() without forced GC

Scenarios

SaaS Small

mode: both (Host, World)
iterations: 1000
warmup: 100
arraySize: 256
effects: false
runs: 3 (averaged)

Orchestrator

mode: both (Host, World)
iterations: 1000
warmup: 100
arraySize: 1024
effects: true
runs: 3 (averaged)

Extreme

mode: both (Host, World)
iterations: 2000
warmup: 200
arraySize: 2048
effects: true
runs: 1

Latency and Throughput

Numbers are averages across runs (except Extreme, which is a single run).

Scenario	Mode	Avg (ms)	p95 (ms)	p99 (ms)	ops/s
SaaS Small	Host	3.52	4.40	7.11	284.03
SaaS Small	World	3.82	4.71	5.70	260.11
Orchestrator	Host	7.49	8.68	10.57	133.75
Orchestrator	World	9.02	14.08	21.24	112.99
Extreme	Host	8.55	12.46	18.37	116.90
Extreme	World	9.07	10.13	11.19	109.85

Memory Trends (World + Snapshot Retention)

This run samples memory every 50 iterations while retaining snapshots to simulate accumulation.

mode: world
iterations: 2000
warmup: 200
arraySize: 1024
effects: true
memSampleEvery: 50
retained snapshots: 40

Heap and RSS (sampled):

heapUsed: 29.33 MB -> 47.75 MB (peak 51.89 MB, delta +18.42 MB)
rss: 103.80 MB -> 151.91 MB (peak 151.91 MB, delta +48.11 MB)

Notes:

Snapshot retention is artificial pressure. Without retention, growth should flatten.
World uses an in-memory lineage by default, which also grows with history length.

Interpretation

The measured latencies are within a practical range for SaaS backends and agent orchestrators with compute-heavy flows.
World adds governance and lineage overhead, which shows up in tail latency (p95/p99) and memory growth under retention.
For self-improving agents, stability depends on explicit retention policies and persistent storage for history to avoid unbounded in-memory growth.

Reproduce

Prereq: pnpm build (benchmarks use packages/*/dist)

bash

# SaaS small
node scripts/bench-core-host-world.mjs \
  --mode both --iterations 1000 --warmup 100 --array-size 256 --no-effects \
  --json .bench/saas-small.json

# Orchestrator
node scripts/bench-core-host-world.mjs \
  --mode both --iterations 1000 --warmup 100 --array-size 1024 --effects \
  --json .bench/orchestrator.json

# Extreme
node scripts/bench-core-host-world.mjs \
  --profile extreme --mode both --json .bench/extreme.json

# Memory sampling (World only, with snapshot retention)
node scripts/bench-core-host-world.mjs \
  --mode world --iterations 2000 --warmup 200 --array-size 1024 --effects \
  --mem --mem-sample-every 50 --retain-snapshots \
  --json .bench/world-mem.json

Limitations

Results vary by CPU, Node version, and OS scheduling.
Memory sampling is approximate and does not force GC.
World is using in-memory storage; production deployments should persist lineage.

Performance Report ​

Summary ​

Environment ​

Methodology ​

Scenarios ​

SaaS Small ​

Orchestrator ​

Extreme ​

Latency and Throughput ​

Memory Trends (World + Snapshot Retention) ​

Interpretation ​

Reproduce ​

Limitations ​