ASGE-RR controller helps AI agents finish 10% more workflows by reserving resources for future calls

AI-agent workflows rely on remote calls to models, memory stores, and tools scattered across a network. As execution advances, these dependency calls form an agentic service graph (ASG), but many calls are only revealed at runtime. This unpredictability creates a resource allocation problem: reserving capacity for a currently visible call may starve a higher-value future call. To tackle this, Trond Vatten and Yuming Jiang formalize the challenge as Agentic Service Graph Embedding (ASGE) and present ASGE-RR, an online controller that maps runtime-revealed calls to replicas and network paths while enforcing capacity, cost, and deadline constraints. ASGE-RR's key innovation is revisable reservations—it protects resources for predicted future calls and updates those reservations as new execution information arrives.

In experiments using OpenHands and GPT Researcher workflows executed with gpt-5.6-luna across a controlled Docker testbed and a WAN testbed, the controller showed clear promise. Every evaluated AI-agent task exposed at least one runtime-revealed dependency call that could be steered before connection establishment. Exploiting that control point, ASGE-RR completed up to 10% more workflow value than a rolling-horizon controller and a current-call steering controller on the WAN testbed. The findings point to a new network-control opportunity: by protecting resources for likely future calls, more AI-agent workflows can finish in time, even in small-scale environments. The paper is available on arXiv (2608.06033) and falls under multiagent systems and performance.