🔠hodoscope
Unsupervised, human-in-the-loop trajectory analysis for AI agents. Summarize, embed, and visualize thousands of agent actions to find patterns across models and configurations.
$ pip install hodoscope
# analyze agent traces
$ hodoscope analyze *.eval
# open the interactive explorer
$ hodoscope viz *.hodoscope.json --open
# analyze agent traces
$ hodoscope analyze *.eval
# open the interactive explorer
$ hodoscope viz *.hodoscope.json --open
Interactive explorer
Every agent action is summarized by an LLM and embedded into a shared vector space. Hodoscope projects these actions onto an interactive 2D map for easy exploration.
Compare across setups
Density difference overlays reveal where one model or configuration behaves differently from the rest. Unique clusters of actions point to behaviors worth investigating.
Bring your own traces
Native support for common agent frameworks. For anything else, trajectories can be passed as simple JSONs.
Inspect AI .eval
OpenHands .jsonl
Docent collection
Raw JSON
{
"id": "trajectory-001",
"metadata": {
"model": "gpt-5",
"arbitrary_metadata": "value"
},
"messages": [
{ "role": "user", "content": "..." },
{ "role": "assistant", "content": "..." }
]
}
"id": "trajectory-001",
"metadata": {
"model": "gpt-5",
"arbitrary_metadata": "value"
},
"messages": [
{ "role": "user", "content": "..." },
{ "role": "assistant", "content": "..." }
]
}
Citation
@article{zhong2026hodoscope,
title={Hodoscope: Unsupervised Behavior Discovery in AI Agents},
author={Zhong, Ziqian and Saxena, Shashwat and Raghunathan, Aditi},
year={2026},
url={https://hodoscope.dev/blog/announcement.html}
}
title={Hodoscope: Unsupervised Behavior Discovery in AI Agents},
author={Zhong, Ziqian and Saxena, Shashwat and Raghunathan, Aditi},
year={2026},
url={https://hodoscope.dev/blog/announcement.html}
}