TonyDzi/agent-sessions

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. Readable mirror: https://github.com/tonydzi/agent-sessions.

This revision (2026-11-01)

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

• 346 sessions, 1 095 871 words
• date range: 2026-08-24 .. 2026-10-01
• most frequent tools: Bash, ToolSearch, Write, Read, Edit, mcp__ccd_session_mgmt__set_session_title
• held back by the secret/PII gate at upload: 0

Selection

This revision still comes from the earlier selective policy (a keyword class plus an independent LLM reviewer). From 1 Nov 2026 the policy is all sessions verbatim; the next daily revision built under it replaces this text and these numbers.

Builder counters for this release (dataset_meta.json): rejected before release: class_hold 137, class_mixed 1 743, class_unknown 179, gate_red 7, short 2 342. 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	43
data/2026-09-00.parquet	300
data/2026-10-00.parquet	3

• 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.
• 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. 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.