TonyDzi/agent-sessions

Agent Sessions: real production transcripts of an LLM coding agent

There are thousands of blog posts about AI agents. There are very few open corpora of real, unedited working sessions where an agent actually builds, breaks and repairs things over 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: fleet synchronisation, instrument building, regression hunting, routine scheduling, retrospectives. Not demos, not curated examples, not prompts written for a paper. The agent is wrong a lot, and the record keeps that in.

What is inside

field	type	meaning
session_id	string	stable pseudonymous id, s_ + 8 hex. Not the real session uuid
date	string	YYYY-MM-DD, from the last timestamp inside the transcript
words	int32	word count of text
turns	int32	number of human + assistant text turns
tools_used	list[string]	distinct tool names the agent called in this session
text	string	the rendered session: turns, tool calls and their arguments

Numbers in this release are printed by the builder, not estimated by hand.

• 455 rows, 1 410 709 words, 5295 turns
• date range: 2026-08-24 .. 2026-10-01
• shortest session 816 words, median ~2600, longest ~15 000
• most frequent tools: Bash, ToolSearch, Write, Read, Edit, PowerShell
• most sessions call at least one tool; the rest are reasoning-only

Two files, same content: agent_sessions.jsonl (one JSON object per line) and agent_sessions.parquet (zstd). dataset_meta.json carries the exact build counters, including what was rejected and why.

Language: predominantly Russian prose with English technical vocabulary, shell commands and code. The operator thinks in Russian and codes in English, and the corpus reflects that rather than hiding it.

How it was collected

The raw material is the agent harness's own .jsonl transcript files, 12 192 of them. The pipeline is deterministic, no LLM in the loop, and each stage is a named gate:

1. Maturity. Only sessions older than 7 days, where age is the last timestamp inside the file. File mtime lies here: reapers and session continuations touch the files, so by mtime only 140 of 1824 files looked older than a week.
2. Size. Files under 200 KB are dropped before rendering, and rendered sessions under 800 words are dropped after. A routine cron run is ten words long; it is not a session worth reading.
3. Render. 97% of a raw transcript is machinery (tool results, screenshots, system reminders). Rendering keeps the human turns, the assistant turns and the tool calls with their arguments. One real session went 1409 KB → 42 KB, 327 records → 104.
4. Class. A keyword classifier sorts each session into publish / mixed / hold / unknown. Only publish is here. See exclusions below.
5. Scrub. Two passes, described below.
6. Gate. Fail-closed: 16 patterns for secrets and private paths, plus 11 more for identifiers the first pass cannot see. Any hit and the row is dropped, with the reason counted in dataset_meta.json → gate_reasons. In this build 6 rows were dropped by the gate: 4 for a residual encoded project path, 2 for a third-party email address.

How it was scrubbed

Replacement, not deletion. A reader needs the shape of a path or a key to follow what the agent was doing, so C:\Users\Anton becomes C:\Users\dev and an AWS key becomes AKIA-REDACTED rather than vanishing.

What is replaced:

• Secrets: private keys, Slack / OpenAI / GitHub / AWS token shapes, inline password=/token= assignments.
• Paths: personal home directories and vault paths, in live form (C:\Users\Anton) and in the harness's dash-encoded form (C--Users-Anton, E---CLAUDE-<workdir>). The encoded form is the one that matters: the first version of this pipeline scrubbed only live paths, so a published file printed C:\Users\dev and C--Users-Anton two lines apart and handed the reader its own mapping table.
• Machine names: every host becomes HUB-01 or NODE-0X.
• Session identifiers: 3116 real session uuids were mapped to the same stable pseudonyms used in session_id, so cross-references between sessions still resolve inside the corpus but do not resolve to anything outside it.
• Numeric identifiers: 29 group chat ids and 415 bare 9-13 digit internal ids (message ids, millisecond timestamps) became stable id_… tokens. The chat-id rule is worth a note: the original pattern ended in a lookahead that excluded a following dot, so a chat id at the end of a sentence was skipped — 29 live ids reached a build before an LLM reviewer, not the gate, pointed at one.
• Email addresses: 187 addresses were replaced. The original rule listed the author's own two addresses by name, which meant every other address passed through untouched. Now any address outside a short allowlist of the author's own public domain and already-pseudonymous service hosts becomes user_…@example.com. Since the 30 Oct 2026 rule revision every address becomes [EMAIL] in full.
• Mask rule v2 (30 Oct 2026). No first letters survive: a first letter plus context is enough to recognise a person. Surnames (third parties and the author's own surname inside the text) become [ФАМИЛИЯ], e-mail addresses [EMAIL], phone numbers [ТЕЛЕФОН], street addresses, flat numbers and postal codes [АДРЕС], ages ("38 лет", "aged 38") [ВОЗРАСТ]. The earlier v1 masks (P***, i***@a***.io, +1-555-0100) were upgraded in every published file. Workplace, company, city and other quasi-identifiers of third parties are not regex-shaped: an LLM reviewer flags them and such a session is withheld.
• People. The speaker label is a role (**User:**), not a name. 163 name forms, including declensions, map to stable pseudonyms: one person is the same pseudonym in every session, so a recurring character stays recurring. All personal names are replaced, including those of public researchers cited in passing. That breaks a few citations, and it is deliberate: telling "public figure" from "private contact" name-by-name is a judgement call that cannot be automated, and the failure mode is someone's identity leaking.

What is not hidden: the operator publishes this corpus under his own name on purpose. His projects, repositories, working hours, city, methods and conclusions are the subject of the dataset, not sensitive content to be removed.

A regex gate cannot see meaning. So a second layer asks an independent LLM, on a different vendor, to read sampled sessions and look for exactly four things: third parties, third-party organisations as employer or client, money tied to a person or deal, and surviving credentials. Verdicts are in _judge_verdicts.jsonl, with the coverage stated as a fraction, not implied.

What is NOT included, and why

Sessions about hiring, leads, investors and money are excluded entirely — not scrubbed, excluded. The classifier flags any session mentioning vacancies, recruiters, interviews, offers, salaries, CRM, outreach, clients, deals, funds, pitches or valuations, and those sessions never enter the corpus.

The reason is that regexes catch keys and paths, but a person is recognisable from three details even after the name is swapped. A conversation about a specific candidate, lead or deal cannot be made safe by substitution, so it is not attempted.

Rejected by class in this build: 845 mixed (technical work that also touches those topics), 121 hold (squarely those topics), 74 unknown. unknown is treated as hold: the default is to not publish.

This also means the corpus is biased towards infrastructure work. It is a record of an agent maintaining a system, not of an agent doing business.

Known limitations

• Single operator, single harness. One person, one agent product, one fleet. Do not read these as representative of agent use in general.
• Semantic review is sampled, not exhaustive. The deterministic gates cover every row; the LLM reviewer covered 20 rows (4.4%), and the verdicts are in _judge_verdicts.jsonl. Read that number as a warning, not reassurance: in those 20 rows the reviewer found three leak classes the regex gate could not see (unmasked group chat ids, third-party email addresses, third-party GitHub handles). Two are now closed by rules; the third is a policy question stated below. A 4.4% sample that yields three new classes is evidence that more classes exist in the 95.6% not yet reviewed.
• Third-party GitHub handles are still present. Handles and owner/repo paths of other people's repositories (maintainers whose public issues and PRs the agent interacted with) are not pseudonymised. The argument for keeping them is that each such interaction is already public on GitHub under that same handle. The argument against — which we take seriously — is that a retained handle lets a reader collect every mention of that person across the whole corpus, and the surrounding internal commentary was never public. Treat this as an open question, not a settled policy.
• Pseudonyms are not anonymity. Sessions describe real public projects. Someone determined could correlate them with the author's public repositories. The protection here is aimed at third parties and credentials.
• Russian-dominant. English-only tooling will see degraded results.
• The classifier is keyword-based. It is crude on purpose, and it errs towards rejection.

Licence and citation

Fixed version (cite this): DOI 10.5281/zenodo.23712140, Zenodo, October 2026. Moving version: Hugging Face dataset (jsonl + parquet, judged-clean rows only) and this repository.

Text: CC BY 4.0. Attribution: Anton Dziatkovskii.

Canonical home of the human-readable transcripts: https://github.com/tonydzi/agent-sessions (reader: https://tonydzi.github.io/agent-sessions/). A versioned snapshot of this dataset is archived on Zenodo with a DOI; cite the DOI for a fixed version and this repository for the moving one.

Residual identifiability. This corpus is pseudonymised, not anonymous, and we do not claim otherwise. The author publishes under his own name, the sessions describe real public projects, and a determined reader can correlate them with his public repositories. The protection is aimed at third parties, credentials and internal identifiers — not at hiding the author.

Takedown. If you find your own name, handle, address or any identifying detail that survived, open an issue at https://github.com/tonydzi/agent-sessions and it will be removed from the next build and from the copies we control. We cannot recall copies already downloaded by others, and we will not pretend we can.