AgentTrove: How to Use a 1.7M-Trace Agent Dataset in Python

AgentTrove is the largest open dataset of AI agent interaction traces, with 1.7 million examples in ShareGPT format. A new Python tutorial shows how to work with the data efficiently to train your own agents.

What AgentTrove Is

AgentTrove gathers real trajectories from a range of AI agents into a single open resource. Each example shows the full sequence: how the agent reads an instruction, breaks down the task, takes actions, processes results and reflects on its mistakes. This level of detail lets researchers study how tasks get solved and which strategies modern systems use.

The dataset covers agents of different types, from simple rule-based systems to complex multi-step solvers. That diversity matters for understanding how agents work and evolve. The ShareGPT format is compatible with popular training tools, from Hugging Face to specialized LLM frameworks.

Key Features

• Data streaming: load the data in chunks, without having to pull the whole dataset into RAM
• Turn normalization: convert agent interactions to a single standard format for consistent analysis
• Strategy and pattern analysis: built-in tools for extracting commands and studying how tasks get solved
• Filtering of successful traces: keep only examples with a correct solution and discard hopeless attempts
• Export to SFT format: a ready-made dataset for supervised fine-tuning of language models, with no extra preparation

How to Use It in Practice

The Python tutorial published with the dataset walks through working with AgentTrove step by step. The first step is to initialize data streaming, which avoids loading everything into memory. That matters especially at this dataset's size, where a full load could require tens of gigabytes of RAM and would needlessly delay the start of analysis.

The next stage is turn normalization. Agents can interact with a system differently depending on the implementation, and converting to a single format simplifies later analysis and comparison of behavior.

Then come the commands: which actions the agent performed, in what order, how it reacted to errors, when it changed strategy and which typical sequences appear often.

Trajectory analysis reveals deeper patterns: which approaches work most often and lead to success, where typical failures occur, and how the agent adapts to new conditions and unexpected obstacles. This is especially useful for understanding failure modes, the places where systems often get stuck.

The final step is to filter successful examples and export a clean SFT dataset for training your own models, free of noise and erroneous trajectories.

What It Means

AgentTrove significantly lowers the barrier to building your own AI agents. Instead of collecting examples from scratch, researchers and developers can now draw on 1.7 million ready-made trajectories from various domains. That should make it possible to iterate faster when building smarter, more reliable and more efficient agent systems.