Position: Stop Hardcoding Multi-Agent Workflows That General Agents Will Outgrow

Xisen Wang

Keywords: Multi-Agent Systems, LLM Agents, Runtime Coordination, Coordination Debt

TL;DR: Static multi-agent workflows encode coordination against capability snapshots that shift every model generation; the field should invest in runtime coordination infrastructure that adapts as agents improve.

Abstract: The agentic AI ecosystem is making a costly bet: that coordination logic written today will remain valid as base agents improve. We argue this bet is wrong. Static multi-agent workflows hardcode roles, routing, and decomposition against capability snapshots that shift with every model generation, producing coordination debt: systems that work for one model, one task, and one moment, but fail to adapt as agents improve or fail in unanticipated ways. This is not an argument against multi-agent systems. It is an argument that the field is doing multi-agent coordination at the wrong layer. LLM agents differ from traditional software components: their capability boundaries are fuzzy, overlapping, context-dependent, and often undiscoverable before execution, and their failures are semantic, plausible, and retry-resistant. We therefore argue that coordination should move from static application logic to runtime infrastructure: discovery, verification, trust, and re-routing as first-class primitives. We synthesize evidence from major agent frameworks, recent decentralized coordination work, and proof-of-concept experiments showing that retry and fixed fallback collapse under systematic and shifting capability failures, while runtime verification and re-routing recover. We call for multi-agent systems that scale with general agents, not workflows that general agents outgrow.

Track: Regular Paper (9 pages)

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 194