AI
Enterprise agents narrow the accuracy gap between global plans and stepwise execution
AgentOrch compares agents that generate complete workflows in advance with agents that act step by step, using 650 enterprise standard operating procedures and 3,220 runtime variants. Stepwise interaction remains more accurate across every evaluated model, but the gap has narrowed substantially. Incremental construction improves global-planning accuracy by 9.98 percentage points, execution-feedback revision by up to 6.13 points on held-out variants, and targeted training by an average of 12.18 points. The results suggest lower-call workflows are becoming more viable, while feedback remains important.
Editorial summary · Publisher article
What matters
- The benchmark matches tools, environments and evaluation across 650 enterprise procedures.
- Stepwise agents remain more accurate, but the gap has shrunk by about 79% from early models.
- Three interventions improve global orchestration: incremental construction, execution feedback and targeted training.
Context & caveats
- The benchmark does not show that global orchestration has eliminated the reliability advantage of stepwise execution.
Why this made the edition
The paper reports a focused benchmark, matched evaluation conditions and quantified improvements for a central enterprise-agent design tradeoff.
Source material