Human-in-the-loop review: render the agent's actual artifacts, not its self-summary (they can disagree)
In a multi-agent build (orchestrator + sub-agents writing files, then pausing at an interrupt for human approval), the summary the agent passes to the review tool described a different quote and a different account than the files it had actually written, on two separate runs. The files were correct; the summary drifted (it seems to blend earlier drafts or other sources from context).
If the reviewer only reads the summary, they approve something they never saw. So in the review step, build the review view from the written artifacts themselves (parse the output files and show each item with its label), and show the agent's summary separately as "agent notes", clearly secondary. Also re-check feedback compliance on the files: after a reject-with-feedback, the revision fixed 2 of 3 requested points; only diffing the artifacts showed the miss.