Multi-agent workflows stop chasing task score alone: LAWA edits the critical path and cuts makespan on seven benchmarks

Automatic learning of multi-agent workflows gains a second objective: shrinking the total time one full execution takes, the makespan, while trying to retain downstream task performance. Earlier automated methods optimized only downstream performance and ignored execution time; when a workflow is reused across recurring queries, the latency of agent calls accumulates over time. Results across seven benchmarks show makespan dropping substantially while task performance stays competitive.

The approach treats a workflow as a graph and edits that graph directly to remove bottlenecks on the critical path. LAWA unifies a graph generator with this graph-editing procedure, so it can both generate workflows from scratch and refine existing ones. The graph edits are theoretically grounded in targeting critical-path bottlenecks, and both generation and refinement are exercised across seven benchmarks.

This report is based on the summary level and did not read the complete paper. The work targets multi-agent workflows for recurring queries, such as AI customer support and AI cloud troubleshooting, aiming to cut makespan while retaining downstream performance; LAWA supports both de novo generation and refinement of existing workflows.

The material attributes latency to model inference time and service congestion and notes that reuse makes latency accumulate; the summary gives no concrete makespan or task-performance numbers per benchmark, and does not state under what conditions the graph edits trade away performance.