Agent sessions at 0.8MB RAM each: benchmark shows sandbox-free harness is dirt cheap
mitsuhiko · x · 2026-09-24
Nathan Flurry shares a memory benchmark for agent sessions: running the Pi harness on Rivet Actors instead of per-session sandboxes takes 0.8MB of RAM per active session, starts in tens of milliseconds, and hibernates when idle.
The core insight: sandboxes pay the overhead of a full Node.js process per session — hundreds of MB plus seconds of startup. Running the harness inside your own backend amortizes that fixed cost across hundreds of agents per process; each new session is just a few JS objects. The measured graph shows a 223MB baseline for Node.js + libraries, with each additional session costing only 0.82MB (content-dependent).
The benchmark code is open-sourced in rivet-dev/rivet, measuring retained resident memory (not peak) with responses verified against a mock LLM endpoint.
More from coding & agent
- Legora hits $200M ARR and cuts agent credit usage 20% with routing and harness rework — chetanp · 2026-09-25
- Free Bots launches a persistent 3D city where AI agents live, work and trade autonomously — Daniel_Farinax · 2026-09-25
- Dev has Claude paint an image pixel-by-pixel in JavaScript and export it as mp4 — tlakomy · 2026-09-25
- Malicious AI agents steal 600K credit cards, hit 100+ sites with skimmers — ChuckDBrooks · 2026-09-25
- Self-organizing agent teams hit 66.7% across five benchmarks, beating their strongest member — Aneesh Pappu · 2026-09-25
- Antigravity SDK offers plug-and-play support for Ollama, LM Studio, and vLLM backends — rseroter · 2026-09-25