- 0:00–0:06
Research question • Measure one agent first • Single-agent baseline • Quality, cost and time
The Research question badge stays fixed in its lane. One blank agent block slides in, labelled "Single-agent baseline". A dashed arrow leads from it to three empty measure slots that draw in one after another under "Quality, cost and time". The slots stay empty and their dashed edges keep moving: nothing has been measured yet. The illustrative task is an open-ended research question. Before any split, one agent or a fixed workflow runs a small set of real tasks, recording quality, cost per accepted result and time.
- 0:06–0:12
Research question • Read why the baseline falls short • Breadth? • Context overflow? • Missing check?
The same agent block stays in place. Three question tags slide in beside it: "Breadth?", "Context overflow?" and "Missing check?". A cyan pointer sweeps past all three, down, back up and down again, without stopping on any: which cause applies is still open. Splitting helps only when the baseline's failures come from breadth, an overflowing context or a missing independent check. Failures from missing knowledge, vague instructions or a weak model rarely improve with more agents. Which applies here is still open.
- 0:12–0:18
Research question • Split only the independent parts • Sub-questions apart • Independent parts • Run in parallel • Dependent steps • One agent or workflow
The article's figure enters whole: "Independent parts", three separate cards, each with its own arrow under "Run in parallel"; and "Dependent steps", three cards linked end to end inside one tray, "One agent or workflow". Markers leave the three independent cards at the same moment and travel their parallel arrows, twice; "Sub-questions apart" is written beside the arrow tips. The dependent chain does not move. Researching sub-questions that do not need each other's answers is one of the signals the article lists for independent breadth. Steps that depend on each other stay with one agent or a fixed workflow.
- 0:18–0:24
Research question • Keep sources through the hand-offs • Trace ID per delegation • Subagent outputs • Sources attached • Merged result • Check against sources
The article's hand-off figure enters: three "Subagent outputs" cards, each with a cyan source tag ("Sources attached"), arrows converging on the "Merged result" card, which holds three tags, and a dashed line to an empty ring, "Check against sources". One after another, a copy of each tagged card travels its arrow into the merged card; a thin trace line writes under each arrow, labelled "Trace ID per delegation". The ring stays empty and the line to it stays dashed: the check has not run. Most failures happen in the hand-offs between agents. Outputs pass through the orchestrator with sources attached, every delegation is logged with a trace ID, and the merged result is checked against sources before it leaves.
- 0:24–0:30
Research question • Compare against the baseline • Cost per accepted result • Criteria agreed first • Decision pending
"Criteria agreed first" is written first, and a bar draws under it. Two blank result columns draw in beneath the bar, side by side: the baseline (one agent block) and the multi-agent version (an orchestrator with three subagents). Each has three empty rows; the bottom rows, "Cost per accepted result", are joined to their label by dashed leaders and stay empty. A dashed "Decision pending" slot sits between the columns. Nothing is filled in. The multi-agent version runs the same set and is kept only if it wins on measures agreed in advance, including cost per accepted result. Here the comparison has not run and the decision is pending.
An illustrative open-ended research question gets a single-agent baseline first, its failures are read, independent sub-questions are split, hand-offs keep their sources, and the multi-agent version is compared on criteria agreed in advance. No result is shown.
Read the article: Multi-Agent Systems Explained: When Splitting Work Across Agents Pays Off