Halo is an open-source project providing tamper-evident runtime records for AI agents. Developed by Brian Kuan, formerly of Vanta, the tool addresses the challenge of verifying AI agent actions when agents are given access to sensitive data. Traditional audit logs can be edited by vendors, making them unreliable. Halo creates a hash-chained, append-only log of every agent action, such as tool or model calls and data access.

This log is verifiable by anyone, ensuring the integrity of the agent's operations. The project includes a Python implementation with zero runtime dependencies and approximately 4,300 lines of code, along with a TypeScript version. Users can integrate Halo with their agents using a simple Python decorator `trace` or via adapters for frameworks like LangChain and OpenTelemetry.

The system proves the integrity of recorded actions but does not guarantee completeness; a separate "witness" component is needed to detect omissions. Halo aims to provide auditable evidence for compliance frameworks like SOC 2, ISO 27001, and the EU AI Act, allowing companies to demonstrate what their AI agents actually did with data, rather than relying on vendor-controlled logs.