Researchers evaluate complete cyclic subtask graphs—maximally flexible multi-agent architectures for tool-using LLM tasks—finding revisitation helps recovery in some domains but adds substantial coordination cost.

Researchers studied complete cyclic subtask graphs, a multi-agent architecture in which executable subtask nodes are fully connected and a unified state-analysis-and-routing agent selects transitions using natural-language criteria, evaluating task-specific and benchmark-generic versions on TextCraft, ALFWorld, and Finance-Agent benchmarks.

The study found three distinct regimes: ALFWorld showed explicit revisitation supports recovery and exploration; TextCraft, a largely prerequisite-chain domain, often favored simpler forward execution; and Finance-Agent remained bottlenecked by retrieval, grounding, and evidence synthesis rather than workflow flexibility alone.

Shared-win token comparisons showed the added flexibility can be substantially more expensive than a single ReAct agent, suggesting multi-agent revisitation helps in some cases but mainly adds coordination cost in others.