Machine-Optimized State Isn't Enough for Human Review

When you automate a workflow, it’s tempting to encode results in whatever form the machine consumes most easily — flags, tags, enum labels. These are great for triggering downstream actions, but they make terrible review surfaces. A reviewer can’t audit correctness from a scattered set of tags.

Keep two representations of the same outcome: one optimized for automation, and one optimized for a human to read at a glance. Write the actual computed values — the score, the category, the reasoning, the chosen path — into a visible, inspectable place like a dashboard or a plain table, alongside the tags that drive the pipeline.

The same instinct applies to agent-generated changes. Propose in a human-legible form, gate the write on a person’s approval, and never let an automated step silently mutate the system of record. Visibility is what makes automation trustworthy enough to keep running.