See the World Through Your AI Agent's Eyes

Ever wondered what your AI agent is actually seeing when it interacts with a website or application? You can read the code, you can see the output, but the agent's internal representation of the screen—its "observation"—has always been a bit of a black box. That makes debugging its decisions or understanding its failures surprisingly difficult.

Agentation is a new open-source tool that flips the script. It gives you a real-time, visual window into your AI agent's perception, turning abstract state data into something you can immediately see and comprehend. It’s like strapping a GoPro to your agent and watching the world from its perspective.

What It Does

In simple terms, Agentation is a visualizer for AI agents. When your agent operates, it typically takes an "observation" of its environment (like a webpage screenshot or a DOM state) and processes it to decide on an action. Agentation captures that raw observation data and renders it as a live, updating visual feed in your browser. You're not seeing the actual website; you're seeing the exact pixel data or structured information that the agent is using as its input to make decisions.

Why It's Cool

The magic here is in the clarity it provides. Instead of guessing why your agent clicked the wrong button, you can see if its observation was cropped, mis-rendered, or missing crucial UI elements entirely. It exposes the gap between what you think the agent sees and what it actually sees.

It's particularly clever for debugging agents that work with computer vision (CV) models, where the observation is a screenshot, or for web agents that might be using simplified HTML representations. You can immediately spot observation errors, timing issues, or environmental quirks that are throwing your agent off track. It turns a frustrating debugging session of log-scrolling into a straightforward visual inspection.

How to Try It

The project is on GitHub and is designed to be integrated into your existing agent setup.

1. Head over to the repository: github.com/benjitaylor/agentation
2. Check out the README for setup instructions. You'll essentially install the package and add a few lines to your agent's code to pipe its observation data to the Agentation server.
3. Run your agent, and open the provided localhost URL in your browser. You'll get a live stream of your agent's observations.

The repo has examples and should get you up and visualizing in a few minutes.

Final Thoughts

As someone who's spent time trying to figure out why an agent suddenly goes haywire, a tool like Agentation feels like a no-brainer for the toolkit. It doesn't change how your agent works; it just gives you the superpower of sight into its process. For developers building, testing, or fine-tuning AI agents—especially for web automation or any visual task—this can save hours of headache. It's a straightforward idea executed to solve a very real, very annoying problem in agent development.

Found an interesting tool? Share it with us @githubprojects.