---
license: cc-by-4.0
language:
  - ru
  - en
task_categories:
  - text-generation
tags:
  - agents
  - llm-agents
  - coding-agents
  - transcripts
  - claude-code
  - agentic-workflows
  - tool-use
pretty_name: Agent Sessions - Real Production Transcripts
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet
---

# Agent Sessions: real production transcripts of an LLM coding agent

Thousands of blog posts describe AI agents; very few open corpora show **real, unedited working
sessions** where an agent builds, breaks and repairs things for months on someone's live
infrastructure. This is one. Every row is one complete session between a human operator and an
LLM coding agent (Claude Code) running on a small fleet of personal machines. Not demos, not
curated examples. The agent is wrong a lot, and the record keeps that in.

The dataset is **revised daily**. Each revision is one commit tagged `vYYYY.MM.DD`; pin a tag or a
commit sha to reproduce a result. Archived snapshots with a DOI: [10.5281/zenodo.23712140](https://doi.org/10.5281/zenodo.23712140).
Readable mirror: [https://github.com/tonydzi/agent-sessions](https://github.com/tonydzi/agent-sessions).

## This revision (2026-11-06)

Numbers below are printed by the release script from the data it uploaded, not typed by hand.

- **12 838 sessions**, **31 765 259 words**
- date range: 2026-08-24 .. 2026-10-30
- most frequent tools: `Bash`, `Read`, `ToolSearch`, `Write`, `Edit`, `PowerShell`
- held back by the secret/PII gate at upload: 0

## Sessions by day

Newest first. Every session id of every day is listed in
[`sessions_by_day.csv`](https://huggingface.co/datasets/TonyDzi/agent-sessions/blob/main/sessions_by_day.csv);
the Dataset Viewer (parquet shards) can sort or filter by the `date` column to read one day.

| day | sessions | words |
|---|---|---|
| 2026-10-30 | 7 | 15 807 |
| 2026-10-29 | 315 | 1 392 867 |
| 2026-10-28 | 238 | 1 084 312 |
| 2026-10-27 | 162 | 553 525 |
| 2026-10-26 | 190 | 350 142 |
| 2026-10-25 | 235 | 581 409 |
| 2026-10-24 | 167 | 781 378 |
| 2026-10-23 | 227 | 1 675 511 |
| 2026-10-22 | 152 | 795 417 |
| 2026-10-21 | 187 | 874 211 |
| 2026-10-20 | 197 | 747 097 |
| 2026-10-19 | 213 | 1 009 920 |
| 2026-10-18 | 136 | 436 414 |
| 2026-10-17 | 131 | 311 638 |
| 2026-10-16 | 163 | 604 687 |
| 2026-10-15 | 182 | 378 853 |
| 2026-10-14 | 263 | 1 526 212 |
| 2026-10-13 | 160 | 333 149 |
| 2026-10-12 | 203 | 405 385 |
| 2026-10-11 | 196 | 315 166 |
| 2026-10-10 | 240 | 184 619 |
| 2026-10-09 | 272 | 474 512 |
| 2026-10-08 | 223 | 677 877 |
| 2026-10-07 | 265 | 1 085 696 |
| 2026-10-06 | 337 | 2 480 463 |
| 2026-10-05 | 134 | 264 411 |
| 2026-10-04 | 125 | 187 557 |
| 2026-10-03 | 180 | 279 817 |
| 2026-10-02 | 260 | 468 450 |
| 2026-10-01 | 257 | 873 315 |
| 2026-09-30 | 269 | 653 538 |
| 2026-09-29 | 249 | 297 409 |
| 2026-09-28 | 245 | 538 787 |
| 2026-09-27 | 219 | 284 648 |
| 2026-09-26 | 292 | 168 265 |
| 2026-09-25 | 239 | 200 861 |
| 2026-09-24 | 556 | 288 906 |
| 2026-09-23 | 489 | 1 041 421 |
| 2026-09-22 | 489 | 500 079 |
| 2026-09-21 | 456 | 649 024 |
| 2026-09-20 | 560 | 475 423 |
| 2026-09-19 | 782 | 198 862 |
| 2026-09-18 | 819 | 215 236 |
| 2026-09-17 | 216 | 187 317 |
| 2026-09-16 | 20 | 116 513 |
| 2026-09-15 | 67 | 552 189 |
| 2026-09-14 | 55 | 274 479 |
| 2026-09-13 | 49 | 184 910 |
| 2026-09-12 | 34 | 44 665 |
| 2026-09-11 | 55 | 336 677 |
| 2026-09-10 | 79 | 545 125 |
| 2026-09-09 | 44 | 263 526 |
| 2026-09-08 | 40 | 101 479 |
| 2026-09-07 | 35 | 129 489 |
| 2026-09-06 | 38 | 99 715 |
| 2026-09-05 | 50 | 278 997 |
| 2026-09-04 | 47 | 498 035 |
| 2026-09-03 | 51 | 348 045 |
| 2026-09-02 | 51 | 183 876 |
| 2026-09-01 | 46 | 432 709 |
| 2026-08-31 | 31 | 52 216 |
| 2026-08-30 | 45 | 166 276 |
| 2026-08-29 | 40 | 86 751 |
| 2026-08-27 | 4 | 19 328 |
| 2026-08-26 | 25 | 52 409 |
| 2026-08-25 | 26 | 79 897 |
| 2026-08-24 | 9 | 68 360 |

## Selection

Since 1 Nov 2026 **every** mature session is released verbatim: work, hiring, money, leads and family included. Topic is no longer a filter. Only two things keep a session out: a secret that cannot be cut, or a personal-data finding that cannot be masked. Such rows are held back and counted, never silently dropped.

Builder counters for this release (`dataset_meta.json`): rejected before release: blurred 45.
A session enters only after it is 7 days old, where age is the *last timestamp inside the
transcript* (file mtime lies: reapers and continuations touch files).

## Files

| file | rows |
|---|---|
| `data/2026-08-00.parquet` | 180 |
| `data/2026-09-00.parquet` | 6 641 |
| `data/2026-10-00.parquet` | 6 017 |

- `data/*.parquet`: zstd shards, one or more per month (`YYYY-MM-NN`). This is what
  `load_dataset("TonyDzi/agent-sessions")` and the Dataset Viewer read.
- `agent_sessions.jsonl`: the same rows as one JSON-lines file, for `grep` and `jq`.
- `sessions_by_day.csv`: `date,session_id,words,turns`, one line per session, newest day first.
- `dataset_meta.json`: exact build counters, rejection reasons, shard list.
- `agent_sessions.parquet` at the repository root is a frozen snapshot from 30 Oct 2026, kept
  for old links; it is not part of the `default` config.

| field | type | meaning |
|---|---|---|
| `session_id` | string | stable pseudonymous id, `s_` + 8 hex; not the real session uuid |
| `date` | string | `YYYY-MM-DD` of the last timestamp in the transcript |
| `words` | int32 | word count of `text` |
| `turns` | int32 | human + assistant text turns |
| `tools_used` | list[string] | distinct tools the agent called |
| `models` | list[string] | model ids seen in the session |
| `effort` | list[string] | reasoning-effort settings seen |
| `text` | string | rendered session: turns, tool calls and their arguments |
| `extra` | string | JSON of any further builder fields, else null |

Language: mostly Russian prose with English technical vocabulary, shell and code.

## How it was rendered

The raw material is the agent harness's own `.jsonl` transcripts. About 97% of a raw file is
machinery (tool results, screenshots, system reminders). Rendering keeps the human turns, the
assistant turns and tool calls with arguments; the speaker label is a role (`**User:**`), not a
name. The pipeline is deterministic code; an LLM reviewer only flags meaning-level personal data.

## Masking (replacement, not deletion)

A reader needs the *shape* of a path or an identifier to follow the agent, so values are
replaced with stable placeholders instead of vanishing.

- **Secrets** (private keys, API/token shapes, `password=`/`token=` assignments, chat ids) are
  cut. A session with a secret that cannot be cut is not released: the gate is fail-closed and
  runs twice, once in the builder and once again in this upload script.
- **Paths and machines**: home directories become `C:\Users\dev`, vault paths a neutral name,
  hosts `HUB-01` / `NODE-0X`, including the harness's dash-encoded path form.
- **Session uuids and long numeric ids** map to stable `s_…` / `id_…` tokens, so references
  between sessions resolve inside the corpus and nowhere else.
- **People** (mask rule v2, 30 Oct 2026): a first name becomes a stable pseudonym, the same in
  every session, so recurring people stay recurring. No first letters survive: surnames become
  `[ФАМИЛИЯ]`, e-mail `[EMAIL]`, phone `[ТЕЛЕФОН]`, street address / flat / postal code
  `[АДРЕС]`, age `[ВОЗРАСТ]`. Public researchers cited in passing are masked too.
- **Quasi-identifiers** (third-party employer, client, city, rare details) are not regex-shaped:
  an independent LLM reviewer flags them and the finding is re-masked (`[ИМЯ]` / `[СКРЫТО]`).
  The reviewer's notes quote what they found, so they are never published.

What is *not* hidden: the operator publishes this corpus under his own name on purpose. His
projects, repositories, methods and conclusions are the subject of the dataset.

If you find something that should have been masked, open a discussion on this repository; the
row is re-masked in the next daily revision.

## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Attribute as: Anton D., *Agent
Sessions: real production transcripts of an LLM coding agent*, Hugging Face, 2026,
doi:10.5281/zenodo.23712140. Tool output quoted inside sessions keeps its upstream license.

## Limitations

One operator, one agent family, one fleet: this is a deep sample, not a broad one. Masking
replaces tokens and so can break the grammar of a sentence. Sessions are rendered, not raw:
tool *results* are dropped, so an agent's claim can be checked only against its next actions.