license: other
license_name: research-only-mixed-upstream
license_link: https://github.com/zzh237/AgentBake
pretty_name: AgentBake Trace Library (100 heterogeneous agents, 7 frameworks)
size_categories:
- 1K<n<10K
tags:
- agents
- personalization
- benchmark
- multi-agent
- traces
AgentBake Trace Library
Execution-trace corpus for the AgentBake benchmark (A Personalization Layer and Benchmark for Heterogeneous Agents). 100 heterogeneous agents spanning seven frameworks (AutoGen, CrewAI, LangChain, LangGraph, LlamaIndex, PydanticAI, Strands), each contributing ~10 recorded multi-turn scenarios.
Contents
good_by_framework/
autogen/<agent>/scenario_XXX/
*__trace_sequence.json # per-step activations: input text, output, tools, timing
*__topology.json # agent graph: nodes, roles, declared capabilities
*__prebuilt_eval.json # LLM-judge scores (coherence, goal success, ...)
selection_metadata.json
crewai/ ... langchain/ ... langgraph/ ... llamaindex/ ... pydanticai/ ... strands/
~4,900 files, ~6.2 GB. 16/14/15/30/5/10/10 agents per framework.
Provenance
Traces were generated by executing rebuilt open-source agents (149-agent
adapter library, pinned upstream repos + commit SHAs in the companion code
repo) on synthetic task scenarios, with Qwen3-32B (qwen.qwen3-32b-v1:0,
via Amazon Bedrock) as the agent backend LLM and as the prebuilt evaluator
judge. The trace content (inputs, outputs, topologies, judge scores) is
generated data; upstream agent code is not included here — see the code
repository for the adapter library and per-repo licenses.
Usage
from huggingface_hub import snapshot_download
snapshot_download("zzh237/agentbake-traces", repo_type="dataset",
local_dir="data/multiagent_traces")
Then run the AgentBake benchmark from the code repo: https://github.com/zzh237/AgentBake
Intended use
Research on agent personalization, orchestration-policy learning, and benchmark evaluation. Scenarios are synthetic; no real user data.