RayTrace
A sandbox tool that records the actual execution of AI coding agents at the kernel level. It runs agents on gVisor and collects execution evidence from the kernel layer, including processes, file systems, and syscalls, providing independent audit ground truth that does not rely on the agent's own account.
Timeline (last 90 days)
RayTrace runs AI coding agents inside gVisor-backed sandboxes and collects execution evidence at the kernel level.
The types of evidence RayTrace records include: every process launched (executable, arguments, working directory, stdout/stderr, exit code, timestamp), the gap between proposed tool calls and actual processes, filesystem diffs, syscall/process tree evidence, and change alerts for protected files.
RayTrace provides a forking a run feature: pause at any step of the agent's execution, change the input or environment the agent sees at that moment, then continue running to observe differences in behavior.
RayTrace's founder describes forking a run as 'the closest thing I've found to a controlled experiment on agent behavior'.
RayTrace is currently in a small-scale pre-pilot state, recruiting 5 Beta testers and offering $10 of free Claude Code sandbox credits in exchange for feedback.
RayTrace's official website is https://raytracer.si.
Kernel-level evidence comes from a layer outside and beneath the agent process, which the agent cannot directly tamper with, and is therefore more trustworthy than application-layer logs.
RayTrace is more likely to appeal to teams that run coding agents frequently in CI/CD or local development, organizations that need security audits or compliance trails for agent behavior, and developers studying agent reliability.
All Facts (8)
Kernel-level evidence comes from a layer outside and beneath the agent process, which the agent cannot directly tamper with, and is therefore more trustworthy than application-layer logs.
50%UnverifiedRayTrace runs AI coding agents inside gVisor-backed sandboxes and collects execution evidence at the kernel level.
50%UnverifiedRayTrace provides a forking a run feature: pause at any step of the agent's execution, change the input or environment the agent sees at that moment, then continue running to observe differences in behavior.
50%UnverifiedRayTrace's founder describes forking a run as 'the closest thing I've found to a controlled experiment on agent behavior'.
50%UnverifiedThe types of evidence RayTrace records include: every process launched (executable, arguments, working directory, stdout/stderr, exit code, timestamp), the gap between proposed tool calls and actual processes, filesystem diffs, syscall/process tree evidence, and change alerts for protected files.
50%UnverifiedRayTrace's official website is https://raytracer.si.
50%UnverifiedRayTrace is more likely to appeal to teams that run coding agents frequently in CI/CD or local development, organizations that need security audits or compliance trails for agent behavior, and developers studying agent reliability.
50%UnverifiedRayTrace is currently in a small-scale pre-pilot state, recruiting 5 Beta testers and offering $10 of free Claude Code sandbox credits in exchange for feedback.
50%