Live agent teams and recorded verification
New agent runs have a Team tab with saved assignments, progress, model usage, file fingerprints and results. Each completed child result is saved immediately, even while a sibling is still working. Recovery reuses those saved results. All children continue to share the parent’s sandbox.
The parent must review the combined result and record relevant check outcomes before completing. For non-code work it can explain why tests do not apply; the view labels that as reviewed. Later changes require another review.
Writes through the file tool are attributed to the child. Shell output files are labeled as reported outputs;
file ownership and conflict prevention are not automatic. Read the same view from Python with run.team().
See shared-sandbox teams.