IETF proposes AI agent affinity framework to optimize CATS traffic steering

The immediate implication is a shift from simple load-balancing to context-aware orchestration, where the network understands the high cost of re-prefilling LLM context. For the streaming and compute ecosystem, this provides a standardized way to manage stateful AI workloads across distributed edge nodes without leaking tenant data or wasting accelerator cycles on redundant loads. By distinguishing between hard locality constraints and soft preferences, operators can finally optimize for latency without violating data residency rules. Watch for the CATS Working Group to integrate these affinity metrics into the formal data model by the April 2027 draft expiry.

The IETF CATS working group has been building a layered set of drafts around computing-aware traffic steering since its charter was approved in 2022. The affinity framework from Y. Mo, D. Yang, and C. Zhou at Huazhong University of Science and Technology is the latest in a series that includes a companion draft on AI agent service characteristics and their implications for Computing-Aware Traffic Steering, which catalogs the workload properties that distinguish AI agent sessions from traditional request-response traffic. Together, these documents signal that the CATS WG is extending its original scope of generic compute-aware routing into the specific demands of stateful, long-running AI inference workloads.

The CATS architecture itself sits within a broader IETF effort to standardize how networks discover and steer traffic toward distributed compute resources. The working group's charter calls for defining requirements, architecture, and protocols for computing-aware traffic steering, and the affinity draft adds a new dimension by formalizing how state locality, storage tier placement, and accelerator readiness should influence service selection decisions. This positions the draft as a building block for operators who need to route AI agent sessions to edge nodes where model context is already loaded, rather than paying the latency and cost penalty of reconstructing that context on every request.

Huazhong University of Science and Technology has been a consistent contributor to the CATS draft series, with the same author team behind multiple related documents. The informational status of this draft means it does not create a standards-track obligation, but it establishes vocabulary and failure-mode taxonomy that implementers and other working groups can reference. If the CATS WG integrates these affinity metrics into its formal data model before the April 2027 expiry, the framework could become the reference point for any vendor building compute-aware steering products that handle AI inference at the edge.