Ask a question that needs research. A supervisor agent splits it into subtasks, researchers gather evidence in parallel, a critic checks the analysis, and a writer turns it into a cited brief. You watch every hand-off.
New runs continue thread and can use its memory.
Each row is an agent; time runs left to right.
The supervisor splits your question into subtasks and decides which ones wait on others.
Researcher agents gather evidence, two at a time. A failed search is retried, then handed to a fresh researcher.
The analyst drafts a comparison. The critic approves it or sends it back with feedback.
The writer turns the approved analysis into a brief with numbered sources.
Replays a past run step by step. Uses no API quota.
Select a step on the timeline.
Open one to see its timeline and brief. Nothing is re-run.
No runs yet.