First-of-its-kind multi-agent observability dataset: 8.9M events, 18 agents, released openly
strickvl · x · 2026-10-15
Developer strickvl, unable to find real multi-agent observability data, built and released zenml/mantiq-swarm-traces on Hugging Face:
- 19 runs of 18 agents collaboratively building a small Lisp interpreter
- A custom recorder captures everything from beneath the agents: programs executed, every file read/write, every model call, all inter-agent messages — 8.9M events across 16 tables
- 10 runs ship with answer keys; 9 are held out as a test set; some runs have planted faults (bad commands, rewritten logs, incomplete specs)
Most existing multi-agent datasets use only 5 agents and none record file I/O, making this the first resource of its kind.
More from coding & agent
- Context windows explained: why bigger isn't the fix for confused agents — Initial_Side3681 · 2026-10-15
- Open-Source Pinloop Uses Claude Code to Auto-Screen Millions of Job Posts — thisdudelikesAI · 2026-10-15
- Kimi Releases Workflow Tutorial #3: Swarm, Plugins and Local Project Sync — Kimi_Moonshot · 2026-10-15
- DataFlex turns LLM training data strategy into four composable, mid-run-adjustable ops — Puzzleheaded_Box2842 · 2026-10-15
- Red-teaming OWASP FinBot: 50 compromises as shipped, 0 with task-scoped enforcement — daniel_tenuo · 2026-10-15
- MiniMax H3 txt2img workflow: native 3072x1536 in 1 minute on a 5060 Ti — -Ellary- · 2026-10-15