AR-FORUM

AR-FORUM / hodoscope

Public

Analyze AI agent trajectories: extract actions, summarize, embed, and visualize.

96
9
100% credibility
Found Feb 21, 2026 at 42 stars 2x -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Hodoscope processes logs from AI agent evaluations, summarizes individual actions using language models, embeds the summaries, and generates interactive 2D visualizations to reveal behavioral patterns across models and tasks.

How It Works

1
📖 Discover Hodoscope

You hear about a helpful tool that turns huge piles of AI agent activity logs into clear pictures of what they're doing.

2
🛠️ Set it up quickly

You add it to your computer with a simple download, and it's ready to go in seconds.

3
🤝 Connect smart helpers

You link a couple of AI services so it can understand and shorten long agent actions into simple descriptions.

4
✨ Feed in your logs

You show it your agent run files, and it chews through thousands of steps, making short notes on each one.

5
📊 Explore the maps

Beautiful interactive charts appear, grouping similar actions into colorful clusters you can zoom and compare.

🎉 Spot the patterns

You easily see how different AI setups behave, find surprises, and get a bird's-eye view of your agents' habits.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 42 to 96 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is hodoscope?

Hodoscope analyzes AI agent trajectories from eval logs, extracting actions, summarizing them with any LLM via LiteLLM (OpenAI, Gemini, etc.), embedding the summaries, and visualizing them as interactive 2D scatterplots. It handles formats like Inspect AI .eval files, OpenHands JSONL, and Docent collections, turning raw logs into cluster maps where behaviors group by model or task. Python CLI commands like `hodoscope analyze run.eval` and `hodoscope viz results/ --group-by model --open` make it dead simple to process thousands of runs.

Why is it gaining traction?

No labels needed—unsupervised t-SNE/UMAP projections plus density overlays spotlight divergences between agent configs, like where one LLM hallucinates tests. Resumable pipelines skip done work, and the single HTML viz packs search, FPS sampling, and group comparisons. Developers digging into agentic AI evals grab it for the quick "aha" on patterns that grep can't reveal.

Who should use this?

AI researchers comparing LLMs on SWE-bench or WebArena evals, teams debugging OpenHands agents, or data scientists deep analyzing agentic large language models for autonomous data science workflows. Perfect if you're sifting Inspect logs to understand failure clusters across models.

Verdict

Promising for ai agent analyze video-style trajectory deep dives, but 1.0% credibility from 12 stars signals beta risks—solid docs and tests mitigate, yet expect rough edges. Install and viz your next eval dump; it'll hook you if agent logs haunt you.

(187 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.