BrowseTrace

A reproducible benchmark for browser-mediated AI agent web workloads.

BrowseTrace is the first public HTTP-request-level benchmark of browser-mediated AI-agent traffic. Unlike agent benchmarks that measure what agents accomplish, BrowseTrace captures how they interact with web infrastructure, enabling cache, CDN, and admission policy research on realistic agentic workloads.

Paper: Anonymous Authors (2026). BrowseTrace: Request-Level Traffic Characterization of Browser-Mediated AI Agents. ACM Internet Measurement Conference (IMC 2026), Karlsruhe, Germany. Author metadata withheld for double-blind review.

Headline results

Metric	Value
Total sessions	1,301
Scripted-random sessions	400 (across 4 regions)
LLM-driven sessions	901 (across 6 models, 5 providers)
Scripted cache-replay requests	82,455
LLM cache-replay requests	357,782
LLM request amplification vs scripted	2–5× (per-task 1–26×)
Rendering overhead	69% of bytes are scripts + stylesheets; 2.9% is HTML

Cache-policy findings at 5 MiB (libCacheSim, reference implementation)

Trace	LRU	GDSF	GDSF advantage
Scripted	37.4%	59.5%	+22 pp
LLM-driven	43.5%	76.2%	+33 pp (amplified on agent traffic)

Full Table 5 (all 6 policies × 5 cache sizes) reproduces from data/traces/full_400_sessions.csv and data/traces/llm_full_901.csv under libCacheSim in seconds.

Six LLMs, five providers

Provider	Model	Sessions
OpenAI	GPT-4.1-mini	350
Google	Gemini 2.5 Flash	150
Google	Gemini 2.5 Pro	150
Anthropic	Claude Haiku 4.5	100
DeepSeek	DeepSeek-V3.2 (agent-trained)	90
Alibaba	Qwen 2.5-Coder 7B (open-weight)	61

Four regions (multi-region cache replay)

Region	Vantage point
Zurich (workstation)	BrowserUse on macOS
US Central	Google Cloud, us-central1 (Iowa)
Europe West	Google Cloud, europe-west1 (Belgium)
Asia Southeast	Google Cloud, asia-southeast1 (Singapore)

Quick start

curl -L -o BrowseTrace.zip https://anonymous.4open.science/api/repo/BrowseTrace/zip
unzip BrowseTrace.zip -d BrowseTrace
cd BrowseTrace
pip install -r requirements.txt

# Reproduce headline cache-policy numbers in <60 seconds
python -c "
from libcachesim import TraceReader, ReaderInitParam, TraceType, LRU, GDSF
p = ReaderInitParam(has_header=True, has_header_set=True, delimiter=',',
                    obj_id_is_num=False, obj_id_is_num_set=True)
p.time_field, p.obj_id_field, p.obj_size_field = 1, 2, 3
for label, path in [
    ('scripted', 'data/traces/full_400_sessions.csv'),
    ('llm',      'data/traces/llm_full_901.csv'),
]:
    for cls, name in [(LRU, 'LRU'), (GDSF, 'GDSF')]:
        r = TraceReader(path, trace_type=TraceType.CSV_TRACE, reader_init_params=p)
        mr, _ = cls(5*1024*1024).process_trace(r)
        print(f'{label:9s} {name:5s} @5MiB: {(1-mr)*100:.1f}%')
"
# Expected:
#   scripted LRU  @5MiB: 37.4%
#   scripted GDSF @5MiB: 59.5%
#   llm      LRU  @5MiB: 43.5%
#   llm      GDSF @5MiB: 76.2%

Release manifest: release-v3

The release-v3 manifest pins four components:

1. Scripted subtree (data/release-v3/) — Zurich scripted-random collection with per-task traces.json, access_log.jsonl, cache_trace.csv, summary.json.
2. Multi-region scripted extension (data/release-v3-geo/) — scripted-random sessions from us-central1, europe-west1, asia-southeast1.
3. LLM-driven bundles — per-model, per-region session directories spanning all six LLMs. See DATASET_CARD.md for full layout.
4. Canonical stitched cache-replay CSVs (data/traces/) — the inputs every paper-level cache claim is replayed from:
   • full_400_sessions.csv (82,455 rows, scripted)
   • llm_full_901.csv (357,782 rows, LLM across 6 models × 4 regions)

The paper compiles every cache-policy number from those two CSVs under libCacheSim. A self-check (verify_submission_gate.py) confirms this on every build.

Trace schema

Each traces.json record (full-fidelity) contains 32 fields organized into four groups: request context (timestamp, URL, method, initiator, resource type, session/task IDs), response metadata (status, content-type, encoded size, protocol, connection reuse), HTTP headers (with Authorization/Cookie/Set-Cookie/Proxy-Authorization stripped per the sanitization policy), and timing (DNS, TLS, TCP, TTFB, transfer in ms).

Cache-replay CSV (cache_trace.csv) is the simulator-ready projection: timestamp_us, cache_key, object_size_bytes, session_id, agent_type.

See schema/trace_schema.py for the canonical schema.

Sanitization policy

All released data has been scrubbed before publication:

• Headers stripped from every request/response: Authorization, Cookie, Set-Cookie, Proxy-Authorization.
• URL query values redacted to _REDACTED_ in traces.json and access_log.jsonl; parameter names are preserved for analytical fidelity.
• User-Agent strings containing project-brand tokens are replaced with BrowseTrace/1.0 (benchmark).
• Cache-replay CSVs preserve full URL uniqueness (required for correct cache keying) but are brand-scrubbed.

Tool: tools/sanitize_release.py. Idempotent; safe to re-run.

Reproducibility

See REPRODUCIBILITY.md for the step-by-step guide from a fresh clone:

• OS requirements, Python version, pip dependencies
• How to reproduce every figure and table in the paper
• How to re-run the submission gate (verify_submission_gate.py)
• How to extend the benchmark with new tasks or models

Extending

Add a new task family or a new LLM:

# Scripted-random baseline for a new task
python collection/runner.py --task my-new-task --surface live \
                            --live-driver scripted-random --repeats 10

# LLM-driven with a specific model
GEMINI_API_KEY=... python collection/runner.py \
    --task all --surface live --live-driver agent --repeats 5 \
    --llm-model gemini-2.5-flash

See collection/tasks.yaml for task definitions and collection/runner.py for the collector.

Citation

If you use BrowseTrace in your research:

@inproceedings{browsetrace2026,
  title     = {{BrowseTrace}: Request-Level Traffic Characterization of Browser-Mediated AI Agents},
  author    = {Anonymous Authors},
  year      = {2026},
  booktitle = {Proceedings of the ACM Internet Measurement Conference (IMC)},
  location  = {Karlsruhe, Germany},
  note      = {Author metadata withheld for double-blind review},
  url       = {https://anonymous.4open.science/r/BrowseTrace},
}

Plain citation: see CITATION.cff.

License

• Code: Apache License 2.0
• Data: Creative Commons Attribution 4.0

Project layout

BrowseTrace/
├── README.md                       — this file
├── CITATION.cff                    — academic citation metadata
├── DATASET_CARD.md                 — dataset provenance, ethics, scope
├── REPRODUCIBILITY.md              — fresh-clone reproduction guide
├── CHANGELOG.md                    — release history
├── LICENSE / LICENSE-DATA          — Apache-2.0 (code) / CC-BY-4.0 (data)
├── requirements.txt                — pinned Python dependencies
├── data/
│   ├── release-v3/                 — Zurich scripted subtree
│   ├── release-v3-geo/             — multi-region scripted
│   └── traces/                     — canonical stitched CSVs
│       ├── full_400_sessions.csv
│       └── llm_full_901.csv
├── tools/
│   └── sanitize_release.py         — scrubber (idempotent)
├── schema/
│   └── trace_schema.py             — canonical trace schema
├── collection/                     — BrowserUse runner, tasks.yaml, Dockerfile
├── analysis/                       — cache simulator + per-section scripts
├── paper/                          — LaTeX source, figures, PDF
└── verify_submission_gate.py       — submission / release gate