Veylith Update: Making AI Agent Actions Safer*
I've been building Veylith, a temporal safety assurance prototype for AI agents, during the hackathon.
The core idea is simple: an action that was authorized a moment ago should not automatically remain safe if the authorization changes before execution. What I've built so far
- A FastAPI backend and interactive web dashboard.
- Deterministic workflow simulations and temporal safety verification.
- SCN-042: Revoked Approval Race — demonstrates an approval being granted, revoked, and followed by an attempted execution.
- Unsafe execution detection, rejection, and explainable counterexamples.
- Evidence-oriented review workflows and regression tests. Recent engineering progress
I've also worked on deployment readiness, correcting the FastAPI entrypoint configuration, improving static asset packaging, and adding an isolated packaging regression test. The coding agent reported 38 passing tests during validation. A successful public deployment is still pending verification. What's next?
The next step is to verify the deployment end to end and continue improving how safety decisions and their supporting evidence are presented to reviewers.
Veylith is a deterministic prototype exploring how temporal verification can help identify unsafe AI-agent actions. It's a step toward making agent decisions more transparent, testable, and accountable.
Explore the project: https://github.com/Rishigit222/Veylith
Log in or sign up for Devpost to join the conversation.