The Scope Creep Automated reviews of new AI-related working group submissions

Review · August 21, 2026

Agent-Awareness in IPv6 Networks (draft-he-dawn-ipv6-agent-aware-framework-00)

draft-he-dawn-ipv6-agent-aware-framework -00View on Datatracker →

Agent-Awareness in IPv6 Networks

Problem Statement

Standard IPv6 forwarding routes packets to reachable destinations. AI agent traffic requires more: a destination must not only be reachable but suitable — possessing the right model family, tool set, or cached execution context. This draft asks whether IPv6 networks should carry in-band "agent-aware metadata" so that cluster-edge routers can select the best agent instance locally, without flooding millisecond-changing state into BGP or IGP.

The motivating pain point is KV-cache affinity: when an LLM inference node already holds a relevant key-value cache for an ongoing session, routing to that node reduces Time-To-First-Token significantly. Current routing is blind to this.

How This Spec Fits In

This is not an OAuth or identity spec — it is a networking / routing proposal submitted to the DAWN (Discover and Actuate Wireless Networks / Distributed AI Networking) working group. It operates at the IP layer, not the application layer.

The spec is Informational and explicitly a discussion starter. It does not define packet formats, extension headers, routing protocol extensions, or IANA allocations. It borrows concepts from:

  • SRv6 / Segment Routing (RFC 8986) — as a candidate carrier for agent-aware metadata
  • In-band Network Telemetry (INT) — as a model for the proposed return-path state-piggyback mechanism
  • Computing-Aware Traffic Steering (CATS) — referenced but not deeply analyzed, despite significant overlap

High-Level Summary

The framework proposes a two-stage forwarding model scoped to operator-controlled domains (data centers, enterprise networks, federation points):

Stage 1 — Macro (Steer to the right cluster): The IPv6 control plane advertises relatively stable information — which clusters offer which capability classes — using a "Semantic Anycast" destination. Normal routing carries packets to the appropriate cluster ingress (an Agent Forwarding Router, AFR). The control plane is not burdened with volatile per-instance state.

Stage 2 — Micro (Pick the best instance): The AFR reads Agent-Aware Metadata carried in the packet, consults a local Agent Forwarding Table (AFT), and forwards to a specific instance within the cluster. The AFT is populated locally via telemetry, instance registration, or a novel "return-path piggyback" mechanism where compute nodes embed updated load and cache-residency signals onto outbound response packets, allowing the AFR to update state at wire speed with zero control-plane overhead.

The metadata carried in-band is categorized into capability hints, policy/trust hints, short-lived state hints, and correlation hints. No specific encoding is proposed; SRv6 SRH, Hop-by-Hop Options, and Destination Options are listed as candidates.

The draft positions itself as complementary to DAWN's discovery work (what exists and how to describe it) and to the Agent Gateway concept (which handles authorization and intent matching at the application layer). The AFR handles only the final local dispatch among already-authorized instances.

Interoperability Assessment

Since this is an explicitly scoped problem statement with no protocol definitions, conventional interoperability evaluation does not apply. That said, a few structural observations are worth noting:

The relationship to CATS is underexplored. The IETF's Computing-Aware Traffic Steering working group has already published RFC 9746 (CATS Framework) and a suite of active drafts that address essentially the same core problem: routing flows to compute instances based on dynamic capability and load information, using a separation between stable service reachability (control plane) and volatile instance state (local scheduling). The draft mentions CATS only in passing as a single reference. Before DAWN invests further in this direction, the authors should do a detailed gap analysis explaining what agent-aware forwarding requires that CATS does not already provide, or propose this work as an extension within CATS rather than a parallel effort.

The metadata categories are vague. Four categories of agent-aware metadata are described at a high level (capability hints, policy hints, state hints, correlation hints), but without format or semantics, it is not possible to evaluate whether the proposed separation of concerns is sound or whether certain categories would leak sensitive information in practice. The authors acknowledge this, but it makes it premature to assess whether the design is correct.

The return-path piggyback mechanism needs validation. The proposal to embed state updates on return packets is elegant but raises practical questions: what happens when return packets are dropped, reordered, or take asymmetric paths? This mechanism is novel enough that it requires both formal analysis and operational evidence before it can be relied upon.

The open questions section is unusually long for a problem statement. Eight substantive questions remain open, including whether L3 agent-awareness is useful at all given application-layer alternatives. This is appropriate for a first draft, but it signals that the design space is far from settled.

Writing Quality

The writing quality is notably good for a -00 individual draft. The problem is motivated with concrete examples (KV-cache affinity, the "reachable but not suitable" scenario), and the scope boundaries are stated clearly and repeatedly. The terminology table is useful and consistently applied.

The document reads more like a position paper than a traditional IETF specification, which is appropriate given its explicitly informational and discussion-oriented nature. Sentences are well-formed and the structure follows IETF conventions. The security and privacy considerations, while brief, correctly identify the main threat classes.

One weakness: the relationship to CATS is the most significant technical context for this work, and it receives only a single-sentence citation. A reader unfamiliar with CATS would not understand whether this draft is an extension, a subset, a replacement, or a divergence. This omission should be addressed before the draft advances.

Author Background

All four authors are from China Unicom and none have published RFCs.

  • Tao He — 0 RFCs published, 5 active drafts, 11 expired drafts. Active across SRv6, routing, and AI/agent networking topics. Most experienced of the group.

  • Mengyao Han — 0 RFCs published, 15 active drafts, 11 expired drafts. The most prolific active writer on the team; work spans routing, agent frameworks, CATS, and transport.

  • Nan Wang — 0 RFCs published, 5 active drafts, 7 expired drafts. Background in DNS, routing, and now agent-aware networking.

  • Tianyi Huang — 0 RFCs published, 4 active drafts, 0 expired drafts. Newer to IETF writing; active on CATS and AI agent topics.

The team is active and writing prolifically across multiple working groups, but no member has yet published an RFC. The breadth of parallel drafts (agent gateways, CATS fallback, ISIS extensions, BGP-LS) from a single organization warrants attention: wide-front participation from a single operator is common in Chinese IETF delegations and may reflect coordinated standards strategy rather than organic community problem-solving. That is not inherently a concern, but reviewers should evaluate whether the problems posed genuinely reflect multi-vendor/multi-operator pain points or are primarily motivated by one organization's deployment architecture.

Overall Assessment

The problem is real: AI agent traffic has characteristics that standard IP routing is blind to, and there is legitimate value in exploring whether and how the network layer can help. The two-stage cluster/instance model is a sensible architectural sketch.

However, this draft should not advance in DAWN without first engaging the CATS working group substantively. CATS RFC 9746 addresses the same structural problem (compute-aware routing with stable capability reachability + dynamic instance selection), and the overlap deserves explicit treatment. If agent-aware forwarding requires mechanisms beyond what CATS provides, that gap should be the focus of a revised problem statement. Alternatively, this work might belong as a CATS use-case or extension document rather than a new DAWN deliverable.