Skip to content

STRIDE: A Systematic Framework for Selecting AI Modalities -- Agentic AI, AI Assistants, or LLM Calls

🕒 Published (v1): 2025-12-01 21:54 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

STRIDE is a design-time framework that scores tasks across four dimensions (task decomposition, reasoning/tool assessment, dynamism attribution, self-reflection requirements) to recommend whether to deploy a direct LLM call, an AI assistant, or fully autonomous agentic AI. Evaluated on 30 enterprise tasks, it achieves 92% modality-selection accuracy while reducing unnecessary agent deployments by 45% and compute costs by 37%.

Problem

Practitioners lack principled criteria for choosing between LLM calls, guided assistants, and autonomous agents at design time, defaulting to intuition-driven over-deployment of agents. This wastes compute, introduces governance risks (uncontrolled tool use, recursive loops), and obscures when true autonomy is warranted.

Method

STRIDE produces an Agentic Suitability Score (ASS) per subtask via a weighted linear combination:

ASS(s) = wr·R(s) + wt·T(s) + ws·S(s) + wρ·ρ(s)

where R = reasoning depth (0–2), T = tool needs (0–2), S = state/memory requirement (0–2), ρ = risk score. Weights are domain-adapted via grid search and refined with RL from deployment outcomes.

A True Dynamism Score (TDS) separates workflow-induced variability from model- or tool-induced variability:

TDS(s) = α·W(s) + β·V(s) − γ·M(s)

A Self-Reflection indicator SR(s)=1 fires when TDS exceeds a threshold and the subtask has conditional branches, nondeterministic tools, or mid-execution validation needs.

Tasks are first decomposed into a DAG of subtasks using a fine-tuned LLM. Aggregated subtask features are passed to a classifier trained on a historical knowledge base to output one of {LLM_CALL, AI_ASSISTANT, AGENTIC_AI}, with persona-aware justifications for developers vs. managers.

Key Contributions

  • First design-time framework for AI modality selection, shifting the decision left before deployment.
  • Quantitative Agentic Suitability Score combining reasoning depth, tool needs, state requirements, and risk, with domain-adaptive weighting.
  • True Dynamism Score that decomposes variability into model-, tool-, and workflow-induced components, gating agentic deployment on the workflow-induced component.
  • Self-reflection requirement assessment as a necessity criterion rather than a performance enhancement.
  • Six-month expert validation with SRE and compliance domain specialists.

Results

  • 92% accuracy in modality selection across 30 tasks (vs. 68% heuristic threshold, 33.3% naive-agent baseline).
  • 45.3% reduction in unnecessary agent deployments (vs. 27.5% heuristic, 0% naive).
  • 37.1% resource savings in compute/API usage (vs. 18.2% heuristic, 0% naive).
  • Domain accuracy: SRE 95%, customer support 93%, compliance 91%, automation 89%.
  • Expert agreement: 78% full, 15% partial, 7% disagreement; 27% improvement in expert alignment over heuristic baseline.
  • Ablation: removing self-reflection caused the largest accuracy drop (92%→76%); removing TDS dropped to 80%.

Limitations

  • Evaluated on only 30 tasks; small scale limits statistical generalizability.
  • Scoring functions are explicitly heuristic, trading interpretability for generality.
  • No multimodal tasks (vision/audio) in current evaluation.
  • Expert validation concentrated in SRE and compliance; other domains lack sustained human calibration.
  • STRIDE sometimes under-estimates dynamism in borderline tasks (e.g., multi-document summarization), recommending assistants when experts prefer agents.

Relevance to Agentic AI / LLM Agents

STRIDE directly addresses the meta-question of agentic AI deployment: when is full autonomy necessary versus wasteful? This is a critical gap as the field moves from benchmarking agent capabilities (SWE-Bench, AgentBench) to governing when agents should be used at all. The TDS's decomposition of variability sources (model vs. tool vs. workflow) provides a principled vocabulary for discussing agent necessity, and the self-reflection gate connects to the broader ReAct/Reflexion line showing that iterative error correction is what distinguishes agents from assistants. For researchers building or evaluating agentic systems, STRIDE offers a complementary design-time filter that could reduce benchmark evaluation surface by pre-selecting tasks that genuinely require agentic treatment.