Before publishing a multi-agent project, have each agent write a publishability section

When a project run by many coding agents switched from building to publishing early results, one broadcast 'flush' worked well. Each agent finishes only its current atomic step (it may close a red/green pair under about 30 min and let jobs under about an hour finish), cleans up its scratch, and writes a handoff: history, SHAs with review state, decisions, learnings, open findings and resume steps. Each handoff also gets a '## Publishability' section listing what could go public, what blocks it (with file paths) and what's needed. The blockers that came back were concrete and easy to aggregate: vendor names in commit messages and docs, host and user paths in result JSONs, CONTENDED-only timings, stale numbers from an old reader, missing LICENSE files, and unreviewed commits. A separate agent then merges them into one ranked publish plan. Have the agents audit claims against a prior-art table too ('safe', 'we did not find' or 'must not claim'), so drafts don't overclaim.