Ingest: A Scheduler-Theoretic Framework for LLM Agent Execution

Type: kb/sources/types/ingest-report.md

Source: scheduler-theoretic-framework-llm-agent-execution.md Captured: 2026-06-28 From: https://arxiv.org/html/2604.11378v1

Classification

Type: design-proposal -- the paper self-describes as "a position paper and design proposal," contributing a theoretical framework, a design analysis, and an experimental protocol, explicitly not a production implementation or empirical results. It carries scientific-paper machinery (formal definitions, theorems, a 70-system survey), but its load-bearing output is an architecture (Graph Harness / SGH) and an untested seven-group evaluation protocol, which puts it in the design-proposal class rather than scientific-paper. Domains: agent-orchestration, scheduling-theory, failure-recovery Author: single author (Hu Wei, sina.com email), no institutional affiliation given. The arXiv id (2604.11378) is future-dated relative to the 2026-06-28 capture; the snapshot itself flags this. Treat as an unreviewed preprint: the formal apparatus is internally checked, but credibility rests on argument quality, not provenance or peer review.

Summary

The paper recasts the dominant "Agent Loop" (ReAct-style observe-reason-act) as a single-ready-unit scheduler: at every step at most one executable unit is active (|𝒰|=1) and the choice of next unit is an opaque LLM inference rather than an inspectable policy. This characterization places Agent Loops and graph executors on one semantic continuum parameterized by ready-set cardinality and policy explicitness/determinism. It then proposes Graph Harness (SGH), which lifts control structure into an explicit static DAG and makes three commitments: plan structure is immutable for the lifetime of a plan version; planning, execution, and recovery are separated into three layers with disjoint execution and diagnostic contexts; and recovery follows a strict three-level escalation (retry → patch → replan) with mechanically enforced no-level-skipping. These trade expressiveness (no competitive/first_of parallelism, no recursive expansion, no parent-chain rollback) for controllability, verifiability, and auditability, with proven bounded termination and a "validation gap" result bounding correctness by per-node validation reliability. No empirical results are reported; a seven-group ablation protocol is specified for future work.

Connections Found

The source maps almost one-to-one onto the KB's clean-model orchestration cluster and is one of the richest source fits in the recent academic-paper batch. Connection discovery surfaced six strong reverse-edge candidates, all in the evidence relationship (the snapshot corroborates existing claims rather than originating them):

  • bounded-context-orchestration-model — SGH's execution tuple ℰ=(𝒮,𝒰,𝒫,𝒪,Δ) is a peer-literature instance of the select/call loop: a symbolic scheduler (𝒫 over ready set 𝒰) driving bounded non-deterministic LLM nodes with results returning to explicit state (Δ). It supplies formal vocabulary (ready-set cardinality |𝒰|) for the note's parallelism remark.
  • scheduler-llm-separation-exploits-an-error-correction-asymmetry — SGH's core move (deterministic topology removes the LLM from policy 𝒫; LLM stays only for node execution/planning/diagnosis) is exactly the bookkeeping/semantic-judgment split the note argues for. The validation-gap theorem (correctness ≥ ∏ p_v, with p_v ≈ 1 for syntactic and ≪ 1 for semantic checks) is independent support that semantic operations resist cheap error correction.
  • enforcement-without-structured-recovery-is-incomplete — SGH's retry → patch → replan ladder with a mechanically enforced escalation invariant is a near-exact instance of the note's corrective → fallback → escalation recovery layer.
  • agent-orchestration-occupies-a-multi-dimensional-design-space — the scheduler continuum (cardinality × policy explicitness × policy determinism) exercises the note's multi-axis claim and offers concrete candidate axes.
  • agent-orchestration-needs-coordination-guarantees-not-just — context partition (𝒞_exec ∩ 𝒞_diag = ∅) and side-effect classification are guarantee primitives layered over the DAG coordination channel.
  • agent-runtimes-decompose-into-scheduler-context-engine-and-execution — SGH's planner/runtime/recovery separation plus context partition is a convergent independent decomposition; the WAL+snapshot persistence maps to the execution substrate.

Four notes were considered and rejected as redundant or loose (agent-is-a-tool-loop, llm-mediated-schedulers-are-a-degraded-variant, the-practical-scheduler-is-the-host-language, topology-isolation-and-verification). All six target notes are currently seedling, so these are load-bearing additive citations, not authority claims — and, per the provenance caveat below, they should be framed as formal/design support, not empirical results.

Extractable Value

  1. Ready-set cardinality |𝒰| as a quantitative orchestration axis -- the KB's design-space note lists scheduler placement, persistence, coordination form, guarantees, and return artifact, but no parallelism/cardinality axis. SGH names exactly this parameter (single- vs multi-ready-unit) and predicts a measurable gain (G_graph) from it, independent of planning quality. This is the highest-reach extraction: a missing dimension for an existing note. [quick-win]
  2. The retry → patch → replan escalation ladder with mechanical enforcement -- a named external instance of the corrective → fallback → escalation recovery layer the enforcement note identifies as missing, complete with a per-node recovery_state counter that rejects level-skipping (attempt_patch requires state ≥ retried; request_replan requires all failed nodes ≥ patched). Concrete operationalization of an otherwise abstract claim. [quick-win]
  3. The validation gap (Theorem 6.3) -- correctness across a passing graph is bounded by ∏ p_v, with syntactic validation p_v ≈ 1 and semantic/LLM-based validation p_v ≪ 1. This is independent formal support for the scheduler-llm-separation note's error-correction asymmetry, and a crisp framing of why semantic checks degrade reliability. [just-a-reference]
  4. Execution/diagnostic context partition (𝒞_exec ∩ 𝒞_diag = ∅) -- isolating failure history from the execution path so failure history cannot leak in as implicit input to subsequent steps. A context-engineering guarantee primitive directly relevant to coordination-guarantees and to the KB's broader context-engineering scope. [experiment]
  5. The scheduler continuum as a comparison lens -- reframes "is a graph executor better than a loop?" as "does moving |𝒰|=1 → |𝒰|≥1 with deterministic 𝒫 help, and by how much?", and classifies LangGraph/TDP/classical workflow engines on shared axes. Reusable vocabulary that improves retrieval and discussion across the orchestration cluster. [just-a-reference]
  6. Empirical-prevalence estimate for parallelism -- the 70-project survey estimates ~30-40% of agent tasks exhibit natural parallelism, the rest being linear chains. This bounds the practical reach of multi-ready-unit scheduling and is a useful (if soft) data point for any claim that parallel orchestration is broadly beneficial. [just-a-reference]
  7. SGH as a whole-system analysis candidate -- the framework is exactly the kind of artifact kb/agentic-systems/ covers (execution loop, orchestration, scheduling, recovery, control surface). A future agentic-systems analysis grounded in this snapshot would be the natural home for the assertions the reverse edges do not absorb. [deep-dive]

Limitations (our opinion)

This is editorial opinion. The dominant limitation is the one the paper itself names first: no experimental validation. Every performance claim (G_graph > 0, G_graph grows with complexity, G_replan helps on failure-prone tasks) is stated as a "logical consequence of the framework's assumptions," which is a way of saying it is unfalsified by data. The seven-group protocol is well-designed on paper but uncalibrated.

Beyond that, the central design claim is hard to vary in the load-bearing direction. The paper concedes that if the planner fails to identify parallel structure, SGH "degenerates to a single-ready-unit scheduler" and its remaining advantages collapse to three (bounded recovery, plan-version immutability, context separation) — none of which require the static DAG that is the paper's headline contribution. So the value of the multi-ready-unit machinery is entirely contingent on planner quality, which the paper does not evaluate and explicitly leaves as an open problem. Combined with the ~30-40% parallelism-prevalence estimate, the honest reading is that the headline mechanism helps a minority of tasks and the durably useful contributions are the recovery protocol and context separation — which the paper itself says are portable even within a plain Agent Loop.

The 70-system survey carries the "controllability-first" and "stable-commitment" principles, but the supporting observations are explicitly labeled qualitative/subjective (e.g., "failure-loop behavior in 3 of 4 graph systems vs 0 of 7 state-machine systems") — small-n and informally coded, so they motivate the design rather than establish it. The LangGraph comparison is fair-disclaimed as structural-not-evaluative, which is appropriate but means the paper does not actually show SGH wins anything against a mature alternative. As a vendor-of-one design proposal, expect favorable framing of the chosen trade-off and missing analysis of the cold-start and implementation-cost burdens (the paper estimates 3,300-6,500 LOC, which it underplays as cheap relative to Airflow). For the KB, the safe use is as a formal articulation of architecture we already model, not as evidence that the architecture performs.

Update agent-orchestration-occupies-a-multi-dimensional-design-space to add ready-set cardinality (|𝒰|, single- vs multi-ready-unit) as an explicit design axis, citing this snapshot as evidence. This is the single highest-reach extraction (it adds a missing dimension to an existing note rather than restating one), the source fit is exact, and it naturally pulls in the scheduler-continuum vocabulary. The remaining five reverse-edge evidence citations and the kb/agentic-systems/ whole-system analysis can follow as separate, lower-priority writes.