Agent Swarms Tested for 10K Hours: Gemini-3.8-Flash Leads SWECraft With 23% Speedup, GPT-6-Astra Worst
mrdrozdov · x · 2026-10-23
World of SWECraft let swarms of frontier agents collaborate for 10,000 hours in simulated agent-owned GitHubs, measuring speedups on real software via the SWECraft Index (held-out benchmarks, CPU instruction counts).
- Gemini-3.8-Flash leads with a geometric mean of 1.232 (23% speedup) across three worlds — the most effective collaborator among all frontier models tested
- Claude-Opus-5 follows at 1.150, then Kimi-K2.7-Code (1.091), GLM-5.3 (1.076), and GPT-5.6-Sol (1.046)
- GPT-6-Astra scores worst among frontier models at just 1.007
- Costs vary wildly: Claude-Opus-5 spent $16,737 while GLM-5.3 hit 1.076 for only $1,283
The study examines what elements enable successful agent collaboration at scale, with direct relevance to multi-agent development workflows.
Related event: Agent swarms collaborate for 10,000 hours in benchmark test(2 posts)→
More from coding & agent
- Solo dev with no coding background ships a playable Godot game in 6 weeks with Opus 5.5 — Delicious-Shower8401 · 2026-10-23
- Agent Standup: open-source live screens for monitoring Claude Code agents — stickygoose · 2026-10-23
- Claude Code's new UI hides activity; switching the transcript view to "Thinking" fixes it — yihui_indie · 2026-10-23
- Morgan Stanley is blocking AI agents — a 6-point playbook for retailers before their checkout is next — sanjaykalra · 2026-10-23
- Connecting Hermes Agent to a local Obsidian vault: read, write, search, link tests — techlatest_net · 2026-10-23
- Your data and evals ARE your agent's intelligence — sarahcat21 · 2026-10-23