Recommender Systems Should Now Be Designed Towards Agents

This position paper argues that recommender systems should now be designed towards agents. We use recommender systems towards agents (RSTA) to denote systems whose immediate consumer is an acting agent, an orchestration layer, or a multi-agent system; whose ranked objects are actionable interventions rather than human-viewable items; and whose success is measured by downstream trajectory utility under preference, cost, policy, and safety constraints. We advance three falsifiable claims: (1) priority-sensitive ranking can improve trajectory utility even when candidate sets are small, (2) service-side information can create value that local planning alone cannot fully recover, and (3) oversight actions such as verify, ask, defer, and escalate should be treated as recommendables rather than post-hoc filters. We sharpen the exclusion boundary against planning, routing, and human-facing recommendation; recast a WorkArena-style hardware-order task family as a full RSTA worked example with an explicit candidate inventory, ranked intervention slate, and trajectory-level objective; and outline an agenda spanning candidate-set reconstruction, oversight-aware ranking, service-to-agent interfaces, multi-agent orchestration, interface-robust evaluation, and governance. The goal is not to relabel all agentic decision making. It is to identify a critical layer: when agents face massive action spaces or bounded compute, ranking dictates which trajectories they can reach—and which catastrophic failures they avoid.