[carlosamg]

those are good points, the way we thought about the "use the latest config" issue was to instead of using references into somewhere else, if the SOP is critical, we ensure it loads those configs in the step, as a deterministic process, that way we know for sure that they were loaded, if something was not loaded the SOP fails loudly and produces an audit log of the failure, so it can be picked up and fixed.

about the snapshot, we are using versioned SOPs so we can keep track and iterate on them, right now if an agent picks a SOP and runs it, it runs the current version, if we improve the SOP the agent should pick up the new one. So the SOP gets loaded as a snapshot, runs once, produces the audit log and ends the run. So the harness won't recheck.

A retry if failed a specific step would be interesting though.

Thank you for your comments!