Source-linked AI summary

STAGE: Stateful Translation to Agentic Graph Execution with Policy-Scoped Context and Deterministic Control

Mengxi Luo, Changjia Chen, An Cao, Zirong Huang, Wanyi Dai

arXiv:2608.22538v1cs.AI

TL;DR

Policy-governed agents must reason over evidence while reliably following complex authorized procedures. STAGE confines model judgment to policy-scoped graph nodes and assigns procedural control to deterministic coordination. Across benchmarks, it generally improves task success and repeated-run reliability, with the largest gains on deeper workflows.

  • Problem

    Long, branching policies require agents to combine semantic reasoning with procedural state tracking, while growing context increases risks of omitted steps, misrouting, and unvalidated judgments.

  • Method

    STAGE uses reviewed executable graphs whose nodes define bounded policy judgments, local context, permitted capabilities, typed results, and coordinator-enforced transitions.

  • Results

    STAGE substantially improves task success and repeated-run reliability across public benchmarks and Smart Dispute, especially on deeper workflows.

  • Takeaways & Limitations

    Policy-scoped context combined with deterministic procedural control improves the reliability of policy execution within the evaluated workflows.

  • Takeaways & Limitations

    STAGE generally involves a model- and workflow-dependent token trade-off, with substantial input-token increases in several public-model settings.

Abstract

from arXiv · show

Policy-governed agents must interpret case evidence while following an authorized procedure. We present \textsc{Stage}, an executable-graph framework that confines model judgment to policy-scoped nodes while placing procedural control in deterministic code. At each node, the model receives task-relevant policy context and returns a typed result, while the coordinator enforces the reviewed execution contract. We evaluate \textsc{Stage} on SOP-Bench Referral Abuse, two $τ^2$-bench domains, and Smart Dispute, a proprietary banking benchmark. Compared with monolithic full-policy execution, \textsc{Stage} generally improves task success and repeated-run reliability across workflows of varying procedural complexity. The largest gains occur on the deeper Telecom and Smart Dispute workflows, where $\mathrm{Pass}^3$ increases by 7.5--55.0 and 57.2--65.7 percentage points, respectively, depending on the model. These results show that combining policy-scoped context with deterministic procedural control can improve the reliability of policy execution.

1 Introduction

Policy-following agents must combine semantic reasoning with reliable execution of long, branching procedures. STAGE addresses this by scoping model judgment to graph nodes while deterministic coordination controls policy execution.

  • Long, branching, multi-stage policies force agents to track both semantic reasoning and procedural state.
  • Growing context increases risks of omitted steps, misrouting, and propagation of unvalidated judgments.
  • STAGE represents policies as explicit action and decision nodes connected by valid transitions.
  • Each node receives only the context, state, and tools required for its local responsibility, while code owns procedural control.
  • STAGE is evaluated against monolithic full-policy execution across public benchmarks and Smart Dispute, with larger gains on deeper, more conditional workflows.

2 Problem Formulation

The study separates graph-owned procedural control from policy localization to test their effects on success, reliability, token use, and robustness to longer execution chains.

  • 2.1 Experimental Conditions: The baseline gives one agent the complete policy and control over both semantic reasoning and procedural progression.
  • 2.1 Experimental Conditions: Graph with full policy uses the reviewed graph for execution while presenting the complete policy at every node.
  • 2.1 Experimental Conditions: Node-scoped STAGE uses the same graph but gives each node only task-relevant policy context grounded in mapped source passages.
  • 2.1 Experimental Conditions: Comparing the baseline with STAGE measures bounded node execution, coordinator-owned control, and localized policy context together.
  • 2.1 Experimental Conditions: Comparing the two graph conditions isolates policy localization while holding control structure fixed.
  • 2.2 Research Questions: RQ1 asks whether graph-structured execution improves task success and repeated-run reliability over monolithic full-policy execution.
  • 2.2 Research Questions: RQ2 tests node-scoped context against complete-policy presentation for success, reliability, and token consumption under fixed control.
  • 2.2 Research Questions: RQ3 examines how performance advantage and error composition vary with execution-chain length.

3 STAGE

STAGE constructs a reviewed, frozen execution graph from authoritative policy and executes bounded policy judgments through contracts validated and routed by a coordinator.

  • An LLM-assisted procedure decomposes authoritative policy into bounded tasks, maps source passages, and authors candidate nodes and transitions.
  • The candidate graph is validated, reviewed, and frozen before evaluation.
  • Each node is an executable contract specifying a local judgment, permitted skills and tools, result validation, and failure behavior.
  • At runtime, the coordinator supplies the node contract, case information, shared memory, and selected policy view to the model.
  • The model returns a typed result with its judgment, supporting case-evidence references, and an outcome when required.
  • The coordinator validates results and enforces graph-declared routing, retries, escalation, termination, and attempt limits.

4 Evaluation

STAGE is evaluated across four workflows with different procedural complexity under controlled comparisons using shared models, inputs, environments, and evaluators.

  • The evaluation covers SOP-Bench Referral Abuse, shallower Retail and deeper Telecom workflows from τ^2-bench, and the deepest Smart Dispute workflow.
  • Public workflows are evaluated with GPT-5.6 Luna, DeepSeek V4 Flash, Claude Haiku 4.5, and Sonnet 5.
  • Smart Dispute is evaluated only with Claude Haiku 4.5 and Sonnet 5 because it contains proprietary banking knowledge.
  • One reviewed and frozen graph is constructed for each workflow before evaluation.
  • Within each comparison, the model, task input, environment, and evaluator remain the same.
  • The protocol compares full-policy and node-scoped STAGE conditions, isolates policy-context effects under fixed graphs, and groups Smart Dispute outcomes by chain length.
  • Pass1 measures single-run success, while Pass3 measures cases completed successfully in all three trials.

5 Results

STAGE generally improves task success and repeated-run reliability across workflows, with especially large gains on deeper Telecom and Smart Dispute procedures. Policy localization also improves reliability and can reduce input-token consumption, while long-horizon completion and runtime efficiency remain constrained.

  • STAGE improves Referral Abuse repeated-run reliability most for weaker models: Luna’s Pass3 rises from 88.0% to 95.5%, and Haiku’s from 79.0% to 88.5%.
  • 32.5 percentage points: DeepSeek’s Retail Pass1 gain is the largest, while Pass3 gains range from 15.0 to 25.0 points across models.
  • Luna’s Telecom Pass1 increases from 47.5% to 90.0%, while Pass3 rises from 32.5% to 87.5%; gains remain visible for all four models.
  • Smart Dispute gains are large: Haiku’s Pass1 rises from 14.3% to 74.3% and Pass3 from 11.4% to 68.6%, while Sonnet’s Pass3 rises from 14.3% to 80.0%.
  • Under fixed graph control, node-scoped context raises Pass3 by 27.5 points for Haiku and 22.5 points for Sonnet, while reducing total token consumption by 37.8% and 56.8%.
  • Smart Dispute gains peak on medium chains at +73.3 points for Haiku and +66.7 for Sonnet; long-chain accuracy declines, with execution/end errors remaining dominant.
  • The current implementation is primarily a reliability and control mechanism rather than a token-reduction method because multiple node-scoped calls can increase output consumption.
  • Luna’s Telecom input consumption decreases by 17.3% while Pass3 increases by 55 points, and Sonnet’s Smart Dispute input consumption decreases by 36.9% alongside a Pass3 increase from 14.3% to 80.0%.

6 Related Work

Prior policy-constrained agents commonly let a persistent agent select the next behavior and use an enforcement layer to judge permissibility. STAGE instead makes a reviewed executable procedure determine the active policy judgment and permitted progression through compliance-carrying nodes.

  • Prior systems such as ShieldAgent and AgentLTL verify actions or traces against temporal rules, while other systems gate tool calls with generated, reviewed, or authored constraints.
  • These approaches share a control boundary in which a persistent agent selects the next task-level behavior and an enforcement layer judges whether it may proceed.
  • Workflow-oriented systems establish explicit workflow state and controller-mediated execution through reviewed SOP DAGs, executable paged programs, phase orchestration, or WCFGs.
  • STAGE centers a compliance-carrying execution node that binds policy passages, case evidence, capabilities, outcomes, schemas, and failure or transition behavior.
  • Within that reviewed interface, the model resolves a bounded semantic question while the coordinator enforces which judgment is active and how the procedure advances.

7 Conclusion

STAGE improves task success and repeated-run reliability by separating local model judgment from graph-owned execution control. Gains persist on long workflows, while remaining longest-chain failures concentrate in execution and terminal completion.

  • STAGE substantially improves task success and repeated-run reliability over monolithic full-policy execution across public benchmarks and Smart Dispute.
  • The Smart Dispute chain-length analysis shows that gains persist on long workflows and remove most errors from skipped or misordered procedural steps.
  • Remaining failures on the longest chains are concentrated in execution and terminal completion.
  • Node-scoped policy context improves success and repeated-run reliability while reducing total context processed relative to full-policy exposure.
  • A conforming trace bounds, reviews, and audits the execution path but does not guarantee a correct semantic judgment.

A Policy-to-graph Construction

The policy-to-graph construction workflow uses LLM-assisted semantic authoring, policy-faithful graph reorganization, and deterministic checks before human review and freezing.

  • LLM-assisted construction segments the policy, extracts procedural tasks, and authors nodes, routing, and shared memory.
  • Candidate fan-in, fan-out, and convergence are reorganized through policy-faithful absorption, serialization, or duplication before rechecking.
  • The resulting plan is deterministically compiled and validated before human review and freezing.
  • Generation stops for review when no faithful graph reorganization is possible, rather than emitting an invented rule.

B Reconstructed Benchmark Graphs

The reconstructed benchmark graphs span compact to deeply conditional workflows, with Retail shallowest, Telecom broader and branched, and Smart Dispute longest and most conditional.

  • SOP-Bench Referral Abuse: SOP-Bench Referral Abuse contains six sequential stages for risk calculation, violation classification, severity assessment, and enforcement.
  • Telecom: Telecom contains 13 nodes and two decisions, adding wider request routing and a deeper technical-support branch.
  • Smart Dispute: Smart Dispute contains 26 nodes and nine decisions, producing the longest and most conditional execution structure.
  • Retail: Retail provides the shallowest routing structure among the reconstructed benchmark graphs.
  • The benchmark graph panels include only non-sensitive node titles and outcome labels.
Loading 2608.22538v1…