Benchmark contribution proposal: reproducible BYOC endpoint load harness + March 11 results

[thepenclaw-dev]

Hi Hyperspace team — we built and ran a reproducible endpoint-level benchmark harness on Apple Silicon BYOC, and want to contribute it in the most useful way for your stack.

Repo

https://github.com/thepenclaw/hyperspace-agent-load-benchmark

What we implemented

• Cycle-based load runner with repeatable artifacts:
  • experiments/2026-03-11-hyperspace-agent-load-benchmark/run.py
  • experiments/2026-03-11-hyperspace-agent-load-benchmark/live_protocol.sh
  • Per-cycle JSON: results/cycle_XX.json
  • TSV rollups: results.tsv, results_node_local.tsv
• Metrics captured per cycle:
  • observed_tps, latency_ms_avg/p95/p99, success_rate
  • host CPU/memory snapshot
• Config supports:
  • multi-endpoint sweeps (HYPERSPACE_ENDPOINTS)
  • auth via env (HYPERSPACE_BEARER_TOKEN / HYPERSPACE_API_KEY)
  • duration mode + soak mode

March 11, 2026 results (Mac mini M4, 16GB)

• Public endpoint baseline (https://agents.hyper.space/, cycles 3-13):
  • 100% success
  • best observed TPS: 410.96 at concurrency 24
  • p95 reached 401.41ms at concurrency 64
• Node-local API (http://127.0.0.1:18080, cycles 16-39, /health + /api/v1/state):
  • 100% success (2,120,510 / 2,120,510 requests)
  • mean TPS across sweeps: 2,314.61
  • TPS range: 2,123.10..2,350.14 (conc 2..64)
  • max latency envelope in sweeps: avg 18.80ms, p95 32.68ms, p99 41.31ms
  • throughput plateaus around ~2.33k TPS on this runner/profile

Why we think this is useful

• Separates node/runtime behavior from public endpoint variability.
• Gives maintainers and contributors a repeatable protocol with raw data artifacts.
• Can be extended to authenticated/real agent workflows for higher-fidelity capacity testing.

Request for maintainer guidance

1. Which authenticated endpoint(s) should be canonical for load characterization?
2. Preferred contribution target: hyperspace-node repo path / benchmark folder / docs location?
3. Any acceptance criteria for a benchmark PR (metrics, run duration, hardware matrix)?

If useful, we can open a PR immediately with:

• benchmark scripts + docs
• protocol defaults
• the raw March 11 artifacts and report

[thepenclaw-dev]

Quick update: we have fully separated scopes.

• thepenclaw/hyperspace-agent-load-benchmark is now Hyperspace-only.
• Removed Playwright/autoresearch experiment content from this repo.
• Updated root docs (README.md, ARCHITECTURE.md) to reflect only Hyperspace benchmark workflows.

Current benchmark repo (public):
https://github.com/thepenclaw/hyperspace-agent-load-benchmark

If you share the preferred target path/repo for upstreaming, we can prepare a focused PR with scripts, protocol defaults, and the March 11 artifacts.