Discussion: Can AI Coding Agents Be Truly Trustworthy Without Verifiable Runtime Environments?

As coding agents become increasingly autonomous, most discussions focus on:
model capability
reasoning quality
tool use
repository access
task completion
However, a deeper question may emerge:
How do we verify the execution environment itself?
For example:
Can developers verify what code was actually executed?
Can runtime behavior be reproduced deterministically?
Can multi-agent coordination be independently verified?
Can execution integrity be proven rather than inferred?
Current approaches often assume trust in:
operating systems
runtime environments
infrastructure layers
external monitoring systems
As AI agents gain greater autonomy and begin operating across repositories, cloud environments, and critical infrastructure, will runtime trust become a first-order systems problem?
Curious how others think about the future relationship between:
AI capability
vs
verifiable execution.
Interested in perspectives from:
AI agent researchers
distributed systems engineers
security researchers
infrastructure architects