Why it matters
This dataset provides invaluable insights into the practical, often messy, realities of AI agent development and deployment. Builders can leverage these real-world examples to understand agent behavior, identify failure modes, and improve the robustness and reliability of their own agent systems.

What changed

The TonyDzi/agent-sessions dataset, available on Hugging Face, comprises a substantial corpus of real-world AI agent interactions. This collection aims to bridge the gap between theoretical descriptions of AI agents and their actual operational performance by including raw conversation logs and research artifacts. The dataset integrates mature Claude Code sessions with exported conversations from ChatGPT and Codex, alongside archived research reports. Unlike typical benchmark datasets, agent-sessions focuses on documenting the authentic, ongoing work of agents, including instances where they err, fail, and attempt self-repair over extended periods on live infrastructure.

The dataset is structured to distinguish between conversation records and research artifacts using provider and artifact_type metadata. Conversation records may contain separately exported sub-agent streams. The dataset is updated via a daily release pipeline, with specific revisions reproducible by pinning a commit SHA. Archived snapshots are also available with a DOI (10.5281/zenodo.23712140), and a readable mirror is hosted on GitHub.

As of the revision dated 2026-11-07, the dataset contains 48,813 records totaling 509,024,270 words, spanning from March 30, 2023, to October 30, 2026. The data originates from live agent workspaces (1 machine, 15,359 files), fleet copies (4 machines, 18,778 files), and a knowledge-vault archive of transcripts dating back to May 2026 (4 machines, 15,654 files). Duplicate sessions are collapsed, resulting in 10,357 unique top-level human sessions. Sessions from other operators are excluded unless they opt-in. Records are distributed across providers: ChatGPT (3,453), Claude (35,095), Codex (8,970), and Deep Research (1,295). The most frequently used tools include Bash, Read, ToolSearch, Write, PowerShell, and Edit. No data was held back by the secret/PII gate at upload.

Newer raw-log builds aim to retain full textual tool inputs and results. However, conversation exports preserve only the text available at the time of export; tool artifacts lost upstream cannot be reconstructed. Opaque binary attachments are withheld pending separate privacy review, with an omission manifest provided. The dataset includes build counters detailing missing, malformed, unsupported, and duplicate sources and records, acknowledging that a complete semantic privacy review is not guaranteed due to deterministic PII scans.

Why it matters for builders

This dataset offers a rare glimpse into the unvarnished reality of AI agent development and operation. By providing access to actual, long-term agent sessions, including their mistakes and recovery processes, it allows builders to move beyond idealized scenarios. Understanding how agents perform in real-world, complex environments, where errors are common and debugging is essential, is critical for developing more reliable and effective AI systems.

Practical impact

Developers can use the TonyDzi/agent-sessions dataset to train and fine-tune their own AI agents, focusing on improving error handling, self-correction mechanisms, and long-term task completion. Analyzing the conversation flows and tool usage patterns can inform the design of more robust agent architectures and better prompt engineering strategies. The dataset's focus on real work, rather than synthetic benchmarks, makes it particularly valuable for building agents intended for practical, open-ended tasks.

Caveats and source limits

The dataset's primary limitation is the potential for incomplete reconstruction of tool artifacts if they were lost upstream before export. While efforts are made to scan for PII, the dataset does not claim complete semantic privacy review. The raw-log builds aim for full textual tool input/output, but older conversation exports may lack this detail. The dataset is continuously updated, and specific revision details are important for reproducibility. The presence of opaque binary attachments requires separate privacy considerations. The dataset is currently available on Hugging Face and GitHub, with a DOI for archived snapshots.

Sources

Written with AI assistance from the linked sources; every claim below was checked against them automatically. How we produce articles.

Claim check: 7/7 supported claims - 7 evidence links - 100% avg confidence
  • The TonyDzi/agent-sessions dataset contains real conversations and research artifacts from AI agents.supported - huggingface.co
  • The dataset includes sessions from Claude Code, ChatGPT, and Codex.supported - huggingface.co
  • The dataset captures actual work sessions, including agent failures and repairs, not just benchmark demonstrations.supported - huggingface.co
  • The dataset contains 48,813 records and 509,024,270 words as of the 2026-11-07 revision.supported - huggingface.co
  • The data spans from March 30, 2023, to October 30, 2026.supported - huggingface.co
  • The dataset is updated daily via a release pipeline, and specific revisions can be reproduced using a commit SHA.supported - huggingface.co
  • The dataset is available on Hugging Face and GitHub, with archived snapshots having a DOI.supported - huggingface.co

Caveats

  • Single-source caution: verify critical details at the linked source.
Radar score 69/100 - how it was calculated
Reliability79
Freshness78
Novelty56
Technical59
Developer78
Ecosystem48
Confidence100
  • Reliability 79: Source trust and evidence quality
  • Freshness 78: Fresh model metadata
  • Novelty 56: Novelty blends source metadata and enrichment
  • Technical 59: Structured technical source signals
  • Developer 78: Model developer utility
  • Ecosystem 48: Single-source caution
  • Confidence 100: Claims have reliable evidence
Share
XLinkedInHacker News

Discussion

Loading comments...

Related articles

AI Coding - Nov 19, 2026Termaxa: AI Agent Shell Command Safety GateTermaxa is a new Rust binary that acts as a gatekeeper for AI coding agents interacting with the shell. It previews command consequences, takes backups, and enforces policies before execution, aiming to prevent accidental damage from agents like Claude Code, Codex, Cursor, and Copilot.Agents - Nov 20, 2026PraisonAI v4.7.26: Offline AI Agents & Enterprise HandoverPraisonAI has released version 4.7.26 of its open-source Python agent framework, enabling fully offline AI agent execution. The update also introduces an early access feature for highly regulated teams, allowing AI tasks to be handed over to self-hosted models with a verifiable audit trail.Agents - Nov 18, 2026Weaxs Stock Analysis Plugin for AI AgentsWeaxs has released a stock research toolkit plugin designed for AI agents, offering 73 tools and 21 skills for market analysis, screening, and backtesting across multiple global markets. This plugin is compatible with various AI agent frameworks, including Pi Agent, Hermes, OpenClaw, and DeepSeek Harness.Agents - Nov 18, 2026Rome OS: Open-Source Agent OS for Recursive AgentsRome OS is an open-source operating system for recursive AI agents, designed to scale the agent's environment rather than just model size. It offers a collaborative, guardrailed space for human-agent interaction, enabling agents to build their own workflows and improve over time.Agents - Nov 18, 2026faf-cli v8.10.0 Enhances AI Agent Context with Project FactsThe faf-cli tool has released version 8.10.0, introducing "Project Facts" to provide AI coding agents with versioned, stack-specific context. This update ensures that AI agents receive accurate information about a project's real stack, reducing guesswork and improving code generation.Agents - Nov 20, 2026SiYuan v3.8.11-rc.3: Knowledge Workspace with AI Agent IntegrationSiYuan has released version v3.8.11-rc.3 of its open-source, privacy-first knowledge workspace. This update enhances the platform's capabilities for human-AI agent collaboration, building on its existing features for personal knowledge management.