SAGA: Workflow-Atomic Scheduling for AI Agent Inference on GPU Clusters

agent 2605.00528 — Cross-paper Synthesis

SAGA (2605.00528) — L3 Per-Paper Synthesis #

SAGA sits at the confluence of two strong 2025–2026 lineages: **program-level agent

scheduling (Autellix) and workflow-aware KV-cache retention** (Continuum, TokenCake,

KVFlow-lineage). Its distinctive move is to fold all three axes — distributed scheduling,

tool-call TTL, and provable fairness — under a single "workflow-as-schedulable-unit" thesis,

then anchor the whole thing to a Bélády competitive-ratio bound. The peer cluster below shows

this is less a novel primitive than a unification-plus-formalization of primitives that

each already existed in isolation.


The 8 peers fall into four bands by how directly they collide with SAGA's claims.

Band A — direct predecessors on the same problem (agent-serving scheduling + KV retention).

Band B — mechanism-complementary (KV reuse / speculation SAGA calls "orthogonal").

Band C — orthogonal-but-adjacent system framings.

Band D — scope/vocabulary peer.


2. 本篇 vs 相关论文的 delta — what's new, incremental, contradictory #

Genuinely new in SAGA:

  1. The empirical competitive-ratio bound (1.31× of Bélády). No peer quantifies how close an
  2. online policy gets to the offline optimum. Continuum's TTL is "expected-utility optimal" with

    no competitive ratio [2511.02230]; Autellix explicitly leaves

    program-level competitive ratio unproven and only simulates the gap to SRPT

    [2502.13965]. SAGA's Bélády framing is the cluster's only attempt at an

    optimality anchor for eviction [2605.00528].

  3. Formal fairness at the task level (AFS + Lyapunov drift, Theorem 2). Autellix has
  4. anti-starvation but no formal bound; Continuum, TokenCake, KVCOMM have none. SAGA is alone

    in providing a provable completion-time deviation bound [2605.00528].

  5. Unification under one thesis. Table 11's 4-axis claim (distributed × tool-TTL × fairness ×
  6. competitive-ratio) is real: no single peer checks all four boxes.

    Incremental / overlapping (SAGA is not first):

    • Tool-call TTL retention. Continuum (2511.02230) and TokenCake (2510.18586) both do this,
    • and Continuum's cost-benefit TTL is arguably more principled (explicit reload + queueing

      terms with a memoryfulness factor η [2511.02230]) than SAGA's percentile-of-

      log-normal TTL with linear pressure scaling. SAGA's §4.2 does not cite Continuum's queueing-

      delay decomposition at all.

    • Program-as-unit scheduling. Autellix (2502.13965) predates SAGA on this framing. SAGA's
    • incremental contribution over Autellix is (a) making structure observable via AEG rather

      than inferred from runtime, and (b) adding cache-retention + fairness on top.

    • Migration-based rebalancing. Both SAGA and Autellix reuse Llumnix-style KV migration;
    • Autellix additionally contributes batched swap (18× fewer ops [2502.13965]).

    Potentially contradictory:

    • SAGA's ablation says **session affinity dominates (+96% TCT when removed), WA-LRU eviction only
    • +54%* [2605.00528]. Continuum instead argues the scheduling/queueing*

      bubble is the dominant cost and KV-reload is secondary [2511.02230]. These

      two agree that co-location/scheduling beats raw cache mechanics — but SAGA foregrounds the

      cache competitive-ratio as its "scientific contribution" while its own numbers say co-location

      carries the win. See §3.


    3. 可攻击面 — adversarial rebuttal against specific claims #

    Attack 1 — the "1.31× of Bélády" headline over-claims relative to what carries the win.

    SAGA calls the competitive ratio "the main scientific contribution" [2605.00528],

    yet its own ablation shows removing session affinity costs +96% TCT vs only +54% for removing

    workflow-aware eviction [2605.00528]. The marquee bound describes the

    least load-bearing lever. A skeptic would say the real engineering win is plain session

    co-location — which Autellix already gets from locality-aware routing without any Bélády framing

    [2502.13965].

    Attack 2 — the GPU-centric latency model may be measuring the wrong bottleneck.

    SAGA attributes 6.0× latency inflation and 38% wasted time to KV regeneration

    [2605.00528]. The CPU-centric characterization (2511.00739) shows that for

    tool-heavy agents, tool execution itself (mostly CPU) is up to 88% of E2E latency

    [2511.00739]. If SAGA's SWE-bench agents spend most wall-time inside

    run_test/retrieval CPU tools rather than KV re-prefill, then eliminating KV regeneration

    optimizes a minority of the latency budget — and the 6.0× figure would be workload-specific to

    cache-regeneration-dominated traces.

    矛盾根源: SAGA's instrumentation counts GPU-side time (KV regen, memory util)

    [2605.00528]; 2511.00739 counts end-to-end time including CPU tool stages

    [2511.00739]. Both can be internally correct: SAGA's 38%

    is "of GPU-attributable time," 2511.00739's 88% is "of E2E time," and they overlap only if the

    tool-idle windows are short. SAGA's own Table 1 (P99 web/API = 45s) suggests long tool tails

    exist, which favors 2511.00739's framing for browser agents.

    Attack 3 — Theorem 2 proves less than the abstract implies. The Lyapunov bound formally

    covers only tenants with W_i ≤ C/N; heavy tenants lie outside the hypothesis yet the paper

    reports comparable 99.1% SLO for them [2605.00528]. The empirical 99.2% is

    called merely "directionally consistent" with the bound. So the "provable bounded-deviation

    guarantee" in the abstract is a guarantee for the easy tenant class, validated empirically for

    the hard one. Continuum is more honest here — it claims no formal guarantee at all

    [2511.02230].

    Attack 4 — TTL design is weaker than a cited-adjacent peer. SAGA's TTL = percentile of a

    fitted log-normal, scaled linearly by memory pressure (§4.2). Continuum's TTL explicitly

    maximizes P(hit)×Benefit − Cost, decomposing benefit into reload and out-of-order (queueing)

    terms weighted by η = −Corr(k, N−k) [2511.02230]. SAGA's TTL ignores the

    queueing-delay benefit that Continuum shows is the dominant term. A reviewer would ask why

    SAGA didn't adopt (or at least compare against) Continuum's cost model.

    Attack 5 — the anti-thrashing gating is fragile by construction. SAGA's steal trigger

    T_idle=100ms is shorter than mean migration latency (230ms), requiring three separate guards

    to avoid oscillation [2605.00528]. This is a self-inflicted design

    tension; TokenCake sidesteps it with a single shared pressure-snapshot ordering protocol

    [2510.18586], arguably cleaner.


    4. 生态位 — paradigm-shift positioning, adoption evidence #

    Paradigm position. SAGA is a consolidation paper, not a paradigm shift. The paradigm shift

    — "treat the agent program, not the request, as the schedulable unit" — was already articulated

    by Autellix (2502.13965) [2502.13965] and Parrot (cited, out of scope). SAGA's

    role is to (1) make the program structure explicit as an AEG rather than inferring it at

    runtime, and (2) bolt on the two things Autellix lacked: cross-tool-call cache retention (which

    Continuum/TokenCake independently supplied) and a formal fairness bound.

    Where it advances the field. The Bélády competitive-ratio framing is a genuine conceptual

    contribution: it reframes "how good is our eviction policy?" from a relative comparison ("beats

    LRU") to an absolute one ("within 1.31× of the offline oracle") [2605.00528].

    This is the kind of anchor the surrounding cluster lacks — Autellix's SRPT-gap simulation

    [2502.13965] is the only comparable move, and it's about scheduling not

    eviction.

    Adoption evidence (weak). SAGA is HPDC'26, built on vLLM v0.6.0 + Ray, and claims

    zero-agent-code-modification deployment [2605.00528]. But code is

    未公开 (unlike Continuum, which ships github.com/Hanchenli/vllm-continuum

    [2511.02230], and RouteLLM, which is fully open

    [2406.18665]). This materially weakens reproducibility relative to

    its closest twin. The reliance on framework-exposed AEG hints also depends on an interop channel

    that the protocol survey (2505.02279) shows is still fragmenting

    [2505.02279].

    Deployment niche. Explicitly latency-sensitive interactive serving, accepting ~30% lower peak

    throughput [2605.00528]. This is the opposite trade from Autellix (which

    optimizes throughput at fixed latency, 4–15× [2502.13965]) —

    so the two are complementary, not substitutes: pick SAGA for interactive multi-tenant TCT, pick

    Autellix for batch throughput.


    5. 未探索方向 — hybrid / adaptive directions from the cluster #

    1. SAGA × Continuum: fold the queueing-delay term into WA-LRU's TTL. SAGA's TTL ignores the
    2. out-of-order/queueing benefit that Continuum quantifies via η [2511.02230].

      A hybrid TTL = SAGA's log-normal percentile plus Continuum's memoryfulness-weighted queueing

      term would likely tighten SAGA's own competitive ratio below 1.31×, since Continuum shows

      queueing (not reload) is the dominant cost the current TTL is blind to.

      1. SAGA × Speculative Tool Calls: multiplicative latency reduction. SAGA itself names this as
      2. future work item 6. Speculative Tool Calls (2512.15834) overlaps tool execution with

        generation, capped at <2× by Lemma 1 [2512.15834]; SAGA schedules the known

        work. Prefetching the argmax-successor KV (SAGA §4.3) and speculatively executing that

        successor's likely tool would compound — but only for stateless tools, per 2512.15834's

        constraint.

        1. CPU-aware AEG accounting. SAGA's work_remain (Eq. 9) counts only GPU prefill+decode
        2. [2605.00528]. Given 2511.00739's finding that CPU tool time can be 88% of

          E2E [2511.00739], an AEG node should carry a CPU-time estimate too, so

          AFS urgency reflects true wall-clock remaining work. Nobody in the cluster co-schedules

          GPU-KV retention and CPU-tool admission (COMB/MAS from 2511.00739) jointly.

          1. Cross-context reuse for multi-agent SAGA. SAGA retains within-session KV; KVCOMM
          2. (2510.12872) reuses KV across agent contexts via offset approximation

            [2510.12872]. A multi-agent SAGA (SAGA §11 item 5) where sibling agents

            share large prompt prefixes could layer KVCOMM's anchor-pool reuse under WA-LRU eviction —

            turning "which session's cache to keep" into "which offset-approximated cache to keep."

            1. Adaptive throughput/latency knob instead of a fixed ~30% trade. SAGA's BFS/DFS hybrid is a
            2. static point [2605.00528]; Autellix's PLAS/ATLAS and RouteLLM's α-threshold

              [2406.18665] both expose a tunable operating point. A workload-adaptive

              controller that slides between throughput-optimal (BFS, Autellix-like) and TCT-optimal

              (SAGA-hybrid) per tenant SLO is unexplored and technically straightforward given the shared

              vLLM substrate all three build on.