Agent Compass - Your AI Agent's Truth Graph to diagnose symptoms

Couldn’t be more excited to finally share Agent Compass with the PH community! Our team poured months into making agent debugging actually painless, can’t wait for you all to try it out 🎉

Thanks for checking out our launch! We’re especially looking for feedback on two things:

1. What frameworks you’d like Compass to integrate with first

2. How you’re currently debugging agent failures
   

Drop your thoughts, we’re here all day answering questions.

Now build AI agents without worrying about things breaking in production. We got you!

Check out our research paper https://arxiv.org/pdf/2509.14647. We achieved state-of-the-art results on error detection and categorization!

Debugging has always been the Achilles’ heel for AI agents, and Agent Compass feels like a true breakthrough. Turning scattered traces into a clear root-cause narrative is exactly what the industry has been missing. The “truth graph” and error tree approach is such a smart way to bring order to the chaos.

Love the emphasis on guided fixes and tracking performance across user journeys. This approach to reliability engineering is exactly what many teams need.

This is amazing and highly useful. Great feature @nikhilpareek and @charu_gupta9 !

Hey PH community! Charu here, Co-Founder at Future AGI.

I’ve noticed a common challenge across AI teams: troubleshooting agents is messy and time-consuming. Even minor tweaks in prompts, tools, or data sources can trigger cascading errors, and most evaluation tools only indicate that something went wrong without showing the reason or solution.

This is where Agent Compass comes in. It requires no setup, automatically tracks failures, uncovers their root causes, and suggests actionable solutions. Teams can spot trends across all their agents, integrate insights into tools like Jira or Slack, and soon enable agents to fix issues on their own.

Try Agent Compass -> https://shorturl.at/GdEhJ

SDK -> https://shorturl.at/fuR2F

Research Paper -> https://shorturl.at/NS5Zs

The 99% accuracy claim is impressive - how do you handle edge cases and unexpected user inputs?

This is exactly the kind of tooling we needed for debugging agents without losing our minds. The Truth Graph and root-cause clustering are 🔥, finally feels like observability is catching up to AI complexity.

@nikhilpareek Looks like a good problem to solve, will check it out. Congrats to the team!