FreeInference Agentic Trace

This dataset contains 16 weeks (2026-05-17 to 2026-09-05) of coding and assistant agents interacting with LLMs through the FreeInference gateway: 12,002 agent sessions from 267 accounts and 14 agent harnesses, with 1,186,582 LLM requests and 1,213,347 tool calls.

The dataset accompanies the paper From Requests to Sessions: A Large-Scale Characterization of Human-Driven Agentic Workloads. The code that reads it, replays its prefix cache and reproduces the paper's figures is at HarvardMadSys/freeinference_agentic_trace.

The trace includes request timing and serving latency, models and anonymized providers, token usage, tool definitions and calls, input mutations, and block-level prefixes for cache replay. No message text is released.

Traces

The primary release is under:

release/<week>/traces/<day>.jsonl.zst

Each file is zstd-compressed JSONL with one agent session per line. A session contains its LLM requests in order. Each request records the model invocation and the tool calls produced by the response. If a tool call launches a subagent, the complete subagent session is nested under that call.

A trace has the following structure:

{
  "week": "2026-05-17_2026-05-23",
  "day": "2026-05-17",

  // Session field
  "session_id": "...",
  "user_id": "...",
  "harness": "claude-code",
  "start_ms": 1778977459236,
  "end_ms": 1778984067198,

  // LLM requests in chronological order.
  "requests": [
    {
      // Position in the session and user turn.
      "step": 0,
      "turn_index": 0,

      // What triggered the request and how it connects
      // to the previous request.
      "step_trigger": "user",
      "link_reason": null,

      // Request timing at the gateway.
      "start_ms": 1778977459236,
      "end_ms": 1778977468828,
      "ttft_ms": 7889,

      // Requested model and anonymized upstream provider.
      "model": "glm-5.1",
      "provider": "B",
      "stream": true,

      // Token counts reported by the provider.
      "provider_tokens": {
        "prompt": 52731,
        "completion": 230,
        "cached": 1280
      },

      // Input tokens broken down by message role, tokenized with tiktoken.
      "role_tokens": {
        "system": 1790,
        "user": 1110,
        "assistant": 15916,
        "tool": 25725,
        "tool_definitions": 8638
      },

      // Shape of the input and tools exposed to the model.
      "n_messages": 179,
      "tool_definition_names": [
        "read_file",
        "write_file",
        "terminal",
        "execute_code"
      ],
      "finish_reason": "tool_calls",

      // How the input differs from the previous request.
      // Null for the first request in a session.
      "mutation": null,

      // Input represented as chained 16-token blocks,
      // tokenized with tiktoken.
      "block_ids": [123, 456, 789, "..."],

      // Tool calls produced by the model response.
      "tool_calls": [
        {
          "call_index": 0,
          "name": "terminal",

          // Normalized tool type. Shell calls additionally record
          // the programs found in the command.
          "category": "*Interpreter",
          "programs": ["python"],
          "parse_status": "success",

          // Arguments preserve their structure while sensitive
          // values are replaced with typed placeholders.
          "arguments": [
            [
              "command",
              "python <path> --status"
            ],
            [
              "timeout",
              "<num>"
            ]
          ],

          // Tool outcome and the gap until the next LLM request.
          "outcome": "error",
          "result_tokens": 842,
          "latency_ms": 5272,

          // If this call starts another agent, its complete
          // session is nested here.
          "subagent": null
        }
      ]
    },

    {
      "step": 1,
      "turn_index": 0,
      "step_trigger": "tool",
      "link_reason": "tool_call_id",

      // Later requests contain the same fields.
      // Here the previous input was extended without changing
      // any earlier content.
      "mutation": {
        "transition": "append",
        "cause": null
      },

      "tool_calls": ["..."]
    }
  ]
}

For convenience, processed/ contains Parquet tables derived by parsing and flattening the traces into sessions, requests, and tool calls.

Notes

Notes

• Request inputs are tokenized with o200k_base, not the provider's tokenizer and grouped into 16-token blocks. Inputs are rendered using the following order: tool definitions → system prompt → messages, Partial final blocks are dropped.
• Block IDs are local to each week. When replaying multiple weeks, make them globally unique, e.g., by prefixing them with the week.
• Sessions are reconstructed by linking related requests through repeated tool-call IDs and assistant messages.
• Sessions do not cross week boundaries. A session that continues into the next week is released as a new session.
• Rare tool names and commands used by fewer than 3 accounts are replaced with other or typed placeholders.
• Each week's manifest.json reports sessions excluded from the release and their exclusion reasons.

Usage

DuckDB

Query the derived Parquet tables directly from the Hugging Face Hub:

SELECT
    sessions.harness,
    count(*) AS requests,
    approx_quantile(requests.ttft_ms, 0.5) AS p50_ttft_ms
FROM 'hf://datasets/harvardMadsys/freeinference_agentic_trace/release/*/processed/requests.parquet' AS requests
JOIN 'hf://datasets/harvardMadsys/freeinference_agentic_trace/release/*/processed/sessions.parquet' AS sessions
    USING (session_id)
GROUP BY sessions.harness
ORDER BY requests DESC;

Hugging Face Datasets

from datasets import load_dataset
traces = load_dataset(
    "harvardMadsys/freeinference_agentic_trace",
    "traces",
    split="week_2026_08_30",
)

The derived request and tool-call tables are also available as separate configurations for convenience:

requests = load_dataset(
    "harvardMadsys/freeinference_agentic_trace",
    "requests",
    split="week_2026_08_30",
    streaming=True,
)
tool_calls = load_dataset(
    "harvardMadsys/freeinference_agentic_trace",
    "tool_calls",
    split="week_2026_08_30",
    streaming=True,
)

Download

Download the full dataset:

hf download harvardMadsys/freeinference_agentic_trace \
    --repo-type dataset \
    --local-dir freeinference_trace

Or download one week:

hf download harvardMadsys/freeinference_agentic_trace \
    --repo-type dataset \
    --local-dir freeinference_trace \
    --include "release/2026-08-30_2026-09-05/*"

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

If you use this dataset, please cite the paper below.

Citation

If you use the dataset or artifact in your research, please cite our paper:

@article{freeinference2026requests,
  title={From Requests to Sessions: A Large-Scale Characterization of Human-Driven Agentic Workloads},
  author={Nixon, William and Tian, Muxin and Zheng, Yunjia and Gunawi, Haryadi S. and Yang, Juncheng},
  year={2026}
}