A2A-MetaTrace

A labeled, metadata-only corpus of multi-agent A2A (Agent-to-Agent) workflow traffic. Each row is one wire message reduced to what a passive network observer sees, obs(m) = (src, dst, t, length, direction), with workflow ground-truth labels; message bodies are discarded. The corpus exists to study how much of a pending agent workflow leaks from communication-graph metadata alone, and to evaluate metadata-protecting transports against it.

Links: GitHub | HuggingFace dataset | Paper (arXiv)

Provenance (disclosed). Workflows run over the real a2a-sdk protocol path (Agent Cards, discovery, message/send, SSE) against real official a2a-samples agent servers backed by real language-model calls. The workflow compositions and labels are ours (see DATASHEET.md). This is the honest provenance claim: the protocol path and agent behavior are real; the composition is designed.

Config

config	agent backend	transport	workflows	classes	variants	rows (messages)
default	agents	https	270	9	27	101516

The corpus is captured from official a2a-samples agents composed into multi-agent workflows; transport is HTTPS-direct (the metadata-protecting transport is evaluated analytically; see DATASHEET.md).

Usage

from datasets import load_dataset
import pandas as pd

ds = load_dataset("a2a-metatrace", split="train")   # message-level rows
df = ds.to_pandas()

# reconstruct workflows and their labels by grouping on trace_id
by_wf = df.groupby("trace_id")
labels = by_wf["task_class"].first()

A workflow is the unit an adversary classifies; featurize per trace_id (message counts, length stats, timing, direction n-grams) and recover task_class. Use the variant column for a leave-variant-out split (generalization to unseen compositions).

Regenerating the corpus

The published Parquet is produced by capturing real official a2a-samples agents. To reproduce it end to end:

1. Get the agents. Clone the official samples repo and point the harness at its Python agents directory:

   git clone https://github.com/a2aproject/a2a-samples.git
   export A2A_SAMPLES_DIR=$(pwd)/a2a-samples/samples/python/agents
   

2. Provide model credentials. The sample agents call real models:

   • export OPENAI_API_KEY=... (the OpenAI- and LiteLLM-backed agents), and
   • for the Google-ADK agents, a Vertex project via Application Default Credentials: export GOOGLE_CLOUD_PROJECT=... and gcloud auth application-default login.

   Keys may instead be placed in a local .envrc (export KEY=VALUE lines); see corpus/sample_agents.py for all configuration variables.

3. Install and run (Python 3.13):

   uv sync
   uv run python -m corpus.run_corpus --runs-per-class 30   # writes results/corpus/a2a_metatrace.json
   uv run python scripts/export_hf.py                        # writes data/ + this README
   

See DATASHEET.md for the workflow classes, provenance, and disclosed model substitutions.

What it is for

• Workflow-fingerprinting / traffic analysis on agent interoperation traffic.
• Evaluating metadata-protecting transports for agent interoperation.
• A reproducible, provenance-disclosed alternative to purely synthetic agent-traffic models.

See DATASHEET.md for construction, intended use, and limitations. Regenerate this card and the Parquet files with uv run python scripts/export_hf.py.

Citation

@misc{a2ametatrace,
  title  = {A2A-MetaTrace: a metadata-only corpus of multi-agent A2A workflow traffic},
  author = {Dangol, Bijaya},
  year   = {2026}
}