Abstract The rapid emergence of Large Language Models (LLMs) and autonomous agents is fundamentally changing the nature of network participants. Traditional Internet infrastructure was designed under the assumption that traffic sources are passive executors of external instructions. Autonomous agents, however, are active optimizers: they formulate plans, adapt strategies, compete for resources, and recursively delegate tasks in pursuit of local objectives. As networks evolve from stateless request systems into intent-driven coordination environments, existing infrastructure mechanisms become increasingly insufficient. Protocols such as TCP, service meshes, orchestration platforms, and rate limiters are effective at managing packets, services, and resources, but they do not account for the long-term coordination consequences generated by autonomous optimization behavior. Without additional coordination constraints, large-scale agent ecosystems may experience retry cascades, resource exploitation, recursive delegation storms, topology fragmentation, and ultimately coordination collapse. This paper introduces the Aegis Fabric Protocol (AFP), a distributed coordination protocol designed for autonomous agent networks. AFP introduces a new infrastructure layer, the Consequence Persistence Layer (CPL), which ensures that coordination outcomes continuously influence future coordination opportunities. Rather than relying on centralized governance, global reputation systems, or application-layer alignment mechanisms, AFP embeds persistent consequences directly into runtime coordination decisions. The protocol is built upon three core principles: consequence persistence, topology resistance, and adaptive coordination friction. Through locally maintained sidecar proxies, agents accumulate coordination history and dynamically adjust future routing eligibility, delegation privileges, and interaction costs. This approach allows network stability to emerge from decentralized consequence enforcement rather than centralized control. AFP is designed to operate across both enterprise agent ecosystems and future open agent networks. In managed environments, AFP mitigates coordination overload, tool storms, and resource contention. In open coordination domains, AFP enables long-term cooperation among autonomous participants without requiring a shared trust authority or globally synchronized state. The central thesis of AFP is that future autonomous systems will not fail primarily because of intelligence, but because of unconstrained optimization. By introducing persistent consequences into distributed coordination, AFP proposes a new infrastructure model for maintaining stable cooperation among autonomous optimizers in the post-stateless era.
Muchen He (Thu,) studied this question.