Agentic Shopping Summary

The agentic shopping literature points to a field at an inflection point. Across twenty-two papers from 2025 to 2026, performance falls sharply as tasks become more complex, especially when agents must manage multiple constraints across turns, which suggests current limits in decomposition, planning, and constraint tracking rather than simple knowledge gaps. The research also shows a clear asymmetry in perception: vision-only systems perform poorly in several benchmark settings, while structured semantic representations such as accessibility trees are much more reliable. Adding vision to accessibility-tree agents yields inconsistent results: on some tasks performance improves, on others it degrades, and no benchmark has yet demonstrated a reliable fusion strategy that consistently outperforms structured navigation alone. Personalization is also foundational at the population level, with persona removal causing large drops in session-level performance, even as individual-level preference matching remains weak. Taken together, the literature suggests that the next advances in agentic commerce will come less from larger models alone and more from better retrieval, stronger stateful architectures, more robust constraint handling, and improved safety design.

• Performance drops sharply as shopping tasks become more complex, especially when agents must satisfy multiple constraints across turns.
• Perception is asymmetric: vision-only systems perform poorly in key benchmarks, while structured semantic navigation such as accessibility trees is far more reliable. Adding vision yields mixed results across tasks.
• Personalization matters at the population level, with persona removal producing large declines in session-level performance, even though individual preference matching remains weak.
• The binding constraint is retrieval quality, not reasoning capability alone. The clearest path forward is better retrieval, stateful memory, stronger constraint handling, and improved safety architecture.

What This Means for Brands

For brands, the message is simple: shopping agents are improving fast, but they still do not behave like expert human shoppers. They struggle when a customer request includes several requirements at once, they perform best when product information is structured clearly, and they still need stronger memory, retrieval, and safety controls before they can be trusted with fully autonomous decisions. The most important takeaway is that success in agentic commerce will depend less on who uses the biggest model and more on who has the cleanest product data, the best search and retrieval systems, and the strongest infrastructure for guiding agents through complex buying journeys.

• Shopping agents are getting better, but they still break down when customer requests become complex.
• Clear, structured product data matters more than brand storytelling alone when agents are choosing what to show.
• Bigger AI models by themselves will not solve the problem; better product data and better system design will.
• Brands should prepare now by improving product feeds, structured attributes, retrieval quality, and safety controls.
• The companies that make their catalogs easiest for agents to understand will have an advantage as agent-driven discovery grows.

The agentic shopping research landscape reveals a field at critical inflection. My analysis of twenty-two foundational papers from 2025–2026 establishes three realities that reshape how we approach agentic commerce, and each one carries direct implications for platform operators, researchers, and governance teams alike.

First, performance degrades predictably with task complexity: not because models lack knowledge, but because they lack hierarchical reasoning structures. Second, vision and structured navigation operate asymmetrically: vision-only agents achieve just 4.55% task success on product search with Claude Sonnet 4 (WebMall, Table 3), while the same model using accessibility trees reaches 56%, and adding vision to accessibility-tree agents yields inconsistent results across tasks and benchmarks. Third, personalization is foundational: in CUSTOMER-R1, persona removal degrades session outcome F1 by 19.72 points (Tables 5 and 8), the largest single ablation effect in the corpus.

Although relatively new, the field has matured rapidly from conceptual frameworks to deployed production systems. Agents achieve roughly 30–55% success on realistic shopping tasks depending on benchmark and complexity (DeepShop: 30% holistic for deep research systems; WebMall: 56% for AX-Tree product search; ShoppingBench: 48.2% for GPT-4.1), with simpler structured tasks reaching substantially higher rates (EcomBench: 90%+ on Level 1). Population-level simulation shows 0.64 correlation with real A/B test outcomes (SimGym, validated across 20 storefronts). Yet a persistent 20–50% performance gap remains. Agents systematically fail at safety-aware decision making (below 40% accuracy), multi-constraint reasoning, and individual preference matching (low single-digit product overlap despite high task-level alignment).

The binding constraint is retrieval quality (41% relevance in production, per Cite Before You Speak, Table 2), not reasoning capability alone. Research attention has been disproportionately directed toward LLM improvements; fundamental gains will come from search quality, decomposition architecture, and safety infrastructure.

The twenty-two papers were selected from the 2025–2026 arXiv corpus using three criteria: the paper must include quantitative evaluation of shopping agent performance (not purely theoretical), it must introduce a benchmark, an architectural contribution, or a production deployment (not a minor variant), and it must be citable at the time of analysis. No systematic search protocol was used; this is a purposive sample weighted toward papers that introduce evaluation infrastructure or production evidence. Papers were identified through citation chaining, arXiv keyword monitoring, and practitioner referral. The sample skews toward English-language work on Western and Chinese e-commerce platforms (Amazon, Lazada, Taobao); multilingual and emerging-market work is underrepresented.

The hidden signal extraction methodology is a structured close-reading protocol applied to each paper in two passes. The first pass extracts headline claims, reported metrics, and stated contributions. The second pass examines subordinate clauses, parenthetical disclosures, appendix tables, and footnotes for information that qualifies, contradicts, or contextualizes the main claims. This approach is inherently subject to confirmation bias: an analyst looking for discrepancies between main text and subordinate text will find them. The synthesis that follows should be read with that limitation in mind. Where a hidden signal is reported, the source location (appendix, footnote, parenthetical) is identified so readers can verify independently.

Each of the twenty-two papers has also been translated into a standalone faithfulness-audited intelligence brief in the companion document (Agentic Commerce Research Intelligence: 20 Paper Translations). That document labels every recommendation as Directly Supported, Cautious Inference, or Unsupported based on the source paper's evidence. The synthesis that follows draws on those translations for specific table and section citations; where a claim in this synthesis is inferential rather than directly measured, it is flagged as such.

EcomBench — Model Performance Comparison — 12 systems clustered between 43–65%. ChatGPT-5.1 leads by 1 point over Gemini. The compressed 22-point spread shows no model is breaking from the pack. These are aggregate scores across difficulty levels — what they hide is what happens when complexity increases, which the findings below make explicit.

Shopping agents fail systematically across all models tested on complex shopping tasks. GPT-4.1 achieves 48.2% average success rate on ShoppingBench (Table 1). Claude Sonnet scores 39.0%. Vision-only agents catastrophically fail at 4.55% (WebMall, Table 3). This is not measurement variance; it is a structural characteristic. The ceiling appears consistently across DeepShop, WebMall, ShopSimulator, and ShoppingBench on harder task types, indicating the problem is inherent to the task architecture, not to any single model family. On simpler tasks, performance can be much higher — EcomBench reports over 90% on Level 1 — but as complexity increases, all models converge toward the same band.

Whether frontier models with extended reasoning (GPT-5.2 and beyond) break through this ceiling on complex tasks remains an open question. The prudent assumption: do not test how your product page works exclusively with frontier models and assume that performance generalizes to production-scale inference.

The 30–50% ceiling is not monolithic. Systematic root cause analysis identifies five distinct failure mechanisms, each quantified and each addressable.

ShoppingBench Fig. 2 — Benchmark Construction Pipeline — Three stages: product sampling from four intent types (complexity increasing from Product Finder down to Coupon & Budget), field sampling across product attributes, and GPT-4o user instruction simulation. The downward arrow labeled "Challenging level" is the key visual — the benchmark is explicitly engineered as a complexity ladder. The performance cliff below is baked into the architecture.

Root Cause 1 — Multi-Constraint Reasoning Collapse

ShoppingBench Fig. 1 — Eight Simultaneous Constraints in Action — A single user request: sunscreen (invisible shield tech, SPF 60, cream texture), foundation stick (full coverage, oil control, built-in brush), budget of 103, conditional voucher valid only when total exceeds 83. The agent calls find_product twice, views candidates, then executes Python to verify budget compliance. The python_execute output shows total_price: 112.0, discount: 10.0, final_price: 102.0, success: true. This is not an edge case. This is the Coupon & Budget tier. This is exactly the architecture that collapses at three-plus constraints.

Simple shopping tasks (single attribute) achieve 59.6% success on ShoppingBench. Adding three simultaneous constraints (budget + brand + availability) causes collapse to 30.4%: a 29.2 percentage point absolute drop, representing approximately 49% relative decline (ShoppingBench, Table 1). ShoppingComp confirms the pattern at scale: product retrieval F1 degrades substantially as constraint count increases (ShoppingComp, Section 5.2). Fine-tuned small models with constraint-aware reward shaping show the mechanism is learnable: Qwen3-4B achieves 48.7% ASR, matching GPT-4.1's 48.2% (ShoppingBench, Table 1).

ShoppingBench Fig. 3 — The Full API Tool Surface — The six tools available to the agent: find_product, view_product_information, python_execute, web_search, recommend_product, and terminate. The python_execute output (total_price, discount, final_price, success: true) shows what financial constraint satisfaction actually requires: code execution, not natural language reasoning. The terminate tool requires a deliberate decision to stop — the action agents systematically underpredict. This is the complete tool surface for the task shown above.

Root Cause 2 — Retrieval Quality Ceiling

Production retrieval systems achieve only 41.1% relevance on real-world e-commerce platforms, compared to ideal scenarios where all evidence is relevant in synthetic benchmarks (Cite Before You Speak, Table 2). This single constraint creates a hard ceiling: CGR drops from 96.52% (synthetic) to 83.86% (real world), a 13-point collapse driven primarily by retrieval quality — though the severity varies by prompting strategy: Citation Prompt mitigates the drop (maintaining 95.46% CGR on real-world data, only a 3.2-point decline from 98.65%), while Guided Prompt suffers a larger 12.6-point collapse. ShoppingComp confirms that retrieval remains the binding constraint for all models tested (Section 5.3). Improving retrieval relevance from 41% to 70% could yield significant absolute performance improvement, potentially 20–30%, without changing agent reasoning capability. This is a projection based on the documented relationship between retrieval relevance and downstream accuracy, not a measured outcome.

OptAgent Framework — Genetic Algorithm Query Optimization — OptAgent attacks the retrieval problem from the query side rather than the index side. The user's initial query is treated as a population seed. LLMs generate variant rewrites via mutation and crossover. A multi-agent evaluator scores each variant against retrieved products, stores scores in memory, and a purchase agent selects survivors. Top-scoring queries breed the next generation. 21.98% improvement over baseline without changing the underlying model. Note: the $8 per query cost (6,109 LLM calls per optimization run, Appendix E) is visible in this architecture if you count the generation cycles. Architectural elegance and operational cost are two different things.

AgenticShop Table 6 — Comprehensive Analysis of Curated Products — The Invalid Page column is the buried story. ChatGPT Search sends users to invalid product pages 71.65% of the time. Claude Sonnet 4 Search: 64.44%. Gemini: 92.36%. Perplexity: 87.61%. Nearly two-thirds to nine-tenths of all product references are broken links or hallucinated pages. The Out of Budget column shows every system clustering between 0.65 and 0.74 — tight clustering that signals a category-wide architectural failure, not a model-specific one. No current system reliably aggregates base price, shipping, tax, and discount into an accurate final cost.

Root Cause 4 — Perception Modality Mismatch

WebMall Table 3 — The Definitive Modality Comparison — The headline: Vision-only produces 0.00% completion, precision, recall, and F1 on End to End tasks for both Claude Sonnet 4 and GPT-4.1. Complete failure — not poor performance, zero. AX-Tree alone puts Claude Sonnet 4 at 62.50% on End to End. AX-Tree + Memory: 75.00% CR, 87.50% precision, 84.26% F1. The memory jump is as large as the modality jump on several task types. The fusion paradox: Claude on Specific Product Search improves 56.06% to 60.23% with vision, but on Vague Product Search AX-Tree + Vision drops to 38.06% vs 53.61% for AX-Tree alone. Vision helps on specific tasks and actively hurts on vague ones.

WebMall Table 4 — Task-Category Disaggregation — Three findings Table 3 hides. First: Checkout is solved — GPT-4.1 achieves 100% CR with AX-Tree + Memory; Claude at 87.50%. The last mile works. The journey to it doesn't. Second: Vision zeros out precisely on judgment-dependent tasks (Add to Cart, Cheapest Offer Vague Requirements, End to End all show 0.00% for Claude under Vision). Third: Find Substitutes is Claude's strongest category at 83.33% CR and F1, holding even partial performance under Vision. Semantic understanding of product relationships is where current agents approach human-level competence. No paper in the corpus has built on this finding.

For Claude Sonnet 4 on Specific Product Search, vision-only agents achieve 4.55% completion rate while accessibility-tree agents achieve 56.06% (WebMall, Table 3). The gap is model-dependent — GPT-4.1 shows a smaller difference on the same task — but the directional advantage of structured navigation is consistent across models. However, adding vision to accessibility-tree agents yields inconsistent results across tasks: on Specific Product Search, performance increases from 56.06% to 60.23%, while on other task types, it decreases (WebMall, Table 3). The strongest conclusion is not that vision hurts, but that no reliable fusion strategy has yet been demonstrated to consistently outperform structured navigation alone. DeepShop confirms the grounding problem: vision-based agents using set-of-mark prompts misclassify interactive buttons, fail to segment review sections, and overlook small filtering widgets (DeepShop, Section 5.4, Figure 12).

Root Cause 5 — Safety Awareness Deficit

ShoppingComp Fig. 1 — Four-Panel Leaderboard: The Confidence-Safety Decoupling — Four panels. Bottom right (Safety Pass Rate): human experts at 77.08%, best model at 37.86%, GPT-4o at 12.50%. Every model below 38%. Bottom left (Report Rationale Validity): GPT-5.2 at 91.73%, human experts at 90.96%. Agents are near-human-level at explaining their choices while being catastrophically below human-level at making safe ones. The reasoning sounds good. The decisions are not. This is not a knowledge gap — it is a confidence-safety decoupling, and this figure is the only place in the corpus where it becomes fully visible.

Best-in-class models achieve approximately 35% safety pass rate on shopping decisions. ShoppingComp reports 35.42% (±6.18%) for GPT-5.2 (Table 3). Human experts achieve 77.08%. The variance is as concerning as the mean: over 5% standard deviation indicates non-deterministic, unstable safety responses across runs (ShoppingComp, Section 5.2). No model has been explicitly trained with safety as a primary optimization objective for shopping decisions. This is not a reasoning limitation (expensive reasoning tokens do not improve safety); it is an objective misalignment.

CUSTOMER-R1 Table 3 — The Inversion: Specialization Defeats Scale — Read the three tiers. Proprietary models cluster between 6.78–11.69% on generated action accuracy — Claude 3.5 Sonnet v2 at 11.69% / 18.54% F1. These are not bad models. They are the wrong models for this task. Fine-tuned Qwen2.5-7B with reasoning: 17.26% / 33.86% F1. A 7-billion parameter open-source model beats every proprietary model in this table on session outcome F1. Llama-3.2-3B with reasoning hits 33.99% F1. Qwen2.5-1.5B jumps 219% on action accuracy with reasoning. Llama-3.2-3B jumps 618% on session outcome F1. The competitive hierarchy is not just closed. It is inverted.

The strategic inflection point: Domain-specific fine-tuning outperforms general-purpose scaling by substantial margins on shopping tasks. Qwen3-4B, fine-tuned with shopping behavior data from a 2.5M-product catalog, matches or exceeds GPT-4.1 (a far larger model) at a fraction of the inference cost (ShoppingBench, Table 1). On CUSTOMER-R1, the zero-shot baseline achieves 7.32% next-action accuracy; SFT+RL reaches 39.58%, a 32-point gain (Table 4). On ShopSimulator, SFT+RL training on Qwen3-8B improves multi-turn success from 6.48% to 35.50%, a 29-point gain (ShopSimulator, Table 3).

This represents a structural advantage: smaller, domain-specialized models consistently match or exceed frontier models on shopping-specific tasks across multiple independent benchmarks. Session-level purchase decisions show comparable gains when personas are integrated during training rather than prompting.

CUSTOMER-R1 Fig. 2 — Framework for Simulating User Behavior — The model holds full HTML observation, complete session behavior history, and a persona ($50,000 income, student, INFJ, cares about brand), then predicts both the rationale and the next action simultaneously. The reward structure separates JSON format validity from action correctness. The persona is a demographic archetype, not a specific person. The rationale ("the product seems nice for my cat, I want to add it to cart") is plausible for the profile and may be completely wrong for the actual individual. Additional signal: this rationale was generated by Claude during training — the model may be learning Claude's reasoning patterns, not actual user cognition.

PAARS Fig. 1 — Designed for Populations, Not Persons — The alignment suite (top right) tells the structural story: query prediction gets individual-level accuracy, item selection gets semantic similarity (softer), and session generation has no individual metric at all — only KL divergence, a population distribution comparison. The system is not failing to achieve individual alignment. It is not trying to. The feedback loop propagates improvements back to the population layer, never to the individual layer. The potential applications box names exactly what this is built for: A/B testing, survey tooling, feature launches, marketplace launches. Every application is population-scale.

Personas improve population-level simulation with the real-world correlation documented in Root Cause 3 (SimGym: 0.64 correlation, validated against actual A/B test outcomes across 20 storefronts). This enables substantial speedup on A/B testing: estimated results in under 1 hour versus 2–4 weeks of live traffic. The persona ablation in CUSTOMER-R1, the largest single ablation in the corpus (19.72-point F1 drop, Tables 5 and 8), confirms that population-level conditioning is load-bearing. PAARS confirms the pattern: persona conditioning improves query alignment by approximately 17% relative (PAARS, Section 5, Table 2). KL divergence for query similarity improves from 18.81 to 17.51 (approximately 6.9% improvement).

But individual-level personalization fails. Agents achieve high task-level alignment yet near-zero exact product overlap in controlled studies: Yang et al. (LLM Agent Meets Agentic AI, Abstract) report approximately 2%. They learn what populations do; they do not learn what individuals prefer. This is architectural: aggregation by design loses individual specificity. Scaling individual personalization without hybrid architecture (personas + explicit user feedback + content-based filtering) yields negative ROI.

ShopSimulator Fig. 3 — Error Statistics of Failed Trajectories (Claude-4-Sonnet) — Panel (b) Personalization Errors: 55.82% of personalization failures involve ignoring personal information entirely — not misinterpreting it, ignoring it. 35.71% involve over-applying it. The agent swings between two failure modes with almost no calibrated middle. Panel (a) Multi-turn Action Errors: 45.63% of buy now failures involve no detail confirmation, 31.31% involve purchase after rejection — the agent retries items the user already refused. Panel (c) Shopper Errors: agents add extra intent (3.69%), distort target intent (2.93%), go silent on key goals (0.43%). This is not a neutral executor of user intent. It actively reshapes what the user wants.

Stateless architectures cannot function for shopping agents. The evidence across multiple benchmarks is consistent: SimGym's memory ablation (Tables 2–4) shows correlation collapse from 0.64 to 0.29, 5.38x increase in stuck-in-loop behavior, and goal completion dropping from 90% to 45%. WebMall confirms that accessibility tree navigation vastly outperforms vision-only modality, as documented in Root Cause 4. ShopSimulator demonstrates that multi-turn success halves without session context (Table 3). Constraint tracking across turns prevents redundant search and failure loop detection.

Simple REST API agents that respond to single queries are becoming less and less plausible for commerce. Production agents require stateful multi-turn session management, constraint preservation across turns, and semantic navigation via accessibility trees.

Twenty-two papers. Six categories. One unified signal: shopping agents exhibit predictable failure modes, measurable capability gaps, and addressable architectural constraints. Full faithfulness-audited translations of each paper are available in the companion document. The inventory below organizes the corpus by contribution type.

AgenticShop Figs. 4 & 5 — Radar Charts and Intent Distribution — Figure 4 (radar charts): every system across both panels collapses on Review Sensitivity and Price Sensitivity while bulging on Brand Preferences. Agents retrieve brand signals correctly and mishandle financial constraints and social proof simultaneously. Figure 5 (intent distribution): Claude spends 80% of session time on Alternative Evaluation and only 8% on Purchase Decision. Agent-E: 64% evaluation, 6% decision. Every system is a thorough researcher and a poor closer. The purchase decision phase — the moment that generates revenue — is where every system invests the least.

Ten patterns emerge across these twenty-two papers. Each carries direct implications for system design, research prioritization, and deployment strategy. I name them here because naming forces precision, and precision is what this field currently lacks.

Theme 08

The Autonomy Calibration Problem

Agents require different autonomy levels for different task types and user demographics. The Autonomy Levels Framework (Feng, McDonald, Zhang) establishes the L1–L5 spectrum with a critical insight: autonomy is a design decision separable from capability. The framework identifies rubber-stamping risk: users disengage when autonomy is too high, approving agent actions without genuine review (Section 3.4). The safety data from ShoppingComp makes the governance case: best model achieves approximately 35–38% safety rubric pass rate versus 77.08% human. That approximately 40-point gap makes unsupervised financial decisions untenable at any autonomy level above L3.

AgenticShop Fig. 5 — Intent Distribution and Capability Radar — Intent distribution (bottom): Claude spends 80% of session time in Alternative Evaluation and only 8% on Purchase Decision. Agent-E: 64% evaluation, 6% decision. Agents that spend 6–8% of their time on purchase decisions are not ready for unsupervised financial autonomy. The rubber-stamping risk at L4 is mirrored in the agent's own action distribution. Radar charts (top): every system collapses on Review Sensitivity and Price Sensitivity while performing best on Brand Preferences — the exact inversion of what matters most for safe financial decisions.

Theme 09

The Cost-Performance Trade-off Opacity

Papers rarely disclose computational costs. When disclosed, costs span four orders of magnitude. OptAgent costs $8 per query with 6,109 LLM calls per optimization (Appendix E). WebMall documents the model-level trade-off: Claude Sonnet 4 costs $0.85–$1.42 per task vs. GPT-4.1 at $0.26–$0.34, but Claude achieves higher completion on end-to-end tasks (Table 5). Most papers report accuracy without reporting cost, making cross-system comparison impossible.

WebMall Table 5 + Fig. 2 — Token Usage, Cost, Runtime, and Cost vs. Completion Rate — The scatter plot makes the trade-off visible in a way prose cannot. Vision-only cluster sits bottom right: high cost, low completion. Claude Vision at $1.42 per task delivers 8% completion — paying the most for the worst outcome. Best value: GPT AX-Tree + Memory at $0.34, 55% completion. Adding memory costs 3 cents and delivers an 11-point completion gain. The table adds three findings the scatter hides: Vision causes Claude to take 45.57 average steps vs 26.67 for AX-Tree — 71% more steps to accomplish less. Memory costs almost nothing on GPT (3 cents, steps decrease) and a 35% cost increase on Claude. Runtime: Claude Vision averages 491.6 seconds per task. Over eight minutes. If real-time shopping assistance is the goal, the runtime column may matter more than the cost column.

The Research-Production Gap

Research results and production reality diverge systematically. Four forces drive this gap: evaluation metric divergence (papers use different success criteria), cost constraints (research uses expensive inference; production must optimize), risk aversion (autonomous shopping carries liability), and data distribution shift (real users behave differently than evaluation datasets). The Perplexity adoption study confirms scale: shopping represents 10% of agentic queries, Amazon dominates at 43.2% of shopping environments, and adoption correlates strongly with GDP (r=0.85–0.86) and moderately with education (r=0.75).

Across all analyses, the hidden signal extraction consistently surfaced information that standard deep research analysis missed. The following ten revelations carry the highest strategic weight. Full evidence chains with faithfulness labels (Directly Supported, Cautious Inference, Unsupported) are available in the companion translations.

1. Retrieval Is the Critical Upstream Constraint, Not Reasoning: Real-world retrieval only 41.1% relevant (Cite Before You Speak, Table 2). Retrieval strategy and constraint enforcement remain the binding constraint for all models (ShoppingComp). The projected 20–30% gain from retrieval improvement is inferential, not measured.
2. Human Evaluation Agreement Is Moderate, Limiting Benchmark Validity: Inter-human κ = 0.592 (OptAgent, Appendix D); 50-sample validation in DeepShop. Reported metrics carry ±10–15% confidence intervals due to evaluation noise.
3. Population-Level Simulation Works; Individual-Level Fails: Confirmed independently by Lu et al. (very low overlap), Yang et al. (approximately 2%), and PAARS (individual alignment remains marginal while group alignment improves).
4. Different Model Families Have Systematically Different Failure Modes: OpenAI: false-positive bias. Gemini: false-negative bias (DeepShop, Section 5.4). ShoppingComp confirms: GPT-5.2 favors breadth-first exploration; Gemini favors depth-first (Section 5.3).
5. Personas Provide Approximately 20-Point Session-Level Improvement: CUSTOMER-R1 ablation: 19.72-point session outcome F1 drop (Tables 5, 8). Effect is concentrated on purchase/terminate decisions (19.72 points) vs. next-action accuracy (1.78 points).
6. Vision-Based Agents Fail Entirely on End-to-End Tasks: 0% success for vision-only on checkout/transaction tasks (WebMall, Table 3). HTML-only Agent-E also fails on visual categories (DeepShop, Table 2).
7. RL Without SFT Initialization Leads to Policy Collapse: RL-only model overselects "click" and never predicts input/terminate actions (CUSTOMER-R1, Table 7). ShopSimulator confirms: standalone RL lags SFT by 6.95 points on multi-turn (Section 4.1).
8. Synthetic Rationales Create Train-Test Distribution Mismatch: Claude-generated rationales improve prediction but may introduce stylistic bias, as the model may learn reasoning patterns specific to the rationale generator rather than actual user cognition (See Think Act; CUSTOMER-R1, Section 3.2).
9. Query Complexity Creates Cliff Effects: RAG achieves 16% on medium, 0% on hard in DeepShop (Figure 5b). Tail queries show largest improvements from optimization (OptAgent, Table 1).
10. Agent Research Costs Span Orders of Magnitude: $0.26–$8.00+ per experiment across OptAgent and WebMall. WebMall documents per-task costs: $0.26–$1.42 depending on model and configuration (Table 5).

Architectural Constraints Revealed Through Candor Zones

Five infrastructure-level limitations emerge from subordinate clauses and appendix disclosures:

1. Action Space Limitations: agents cannot interact with sliders, dropdowns, lazy-loaded content (DeepShop, Section 5.4, Figure 14).
2. Context Window Pressure: HTML truncation loses middle history; WEBSERV addresses this with DOM parsing that streamlines raw HTML while preserving interactive elements.
3. Multi-UX-Inference and paged attention architecture for handling feature interference in production (Cite Before You Speak, Section 5).
4. Incus Container Launch (1.78 seconds vs Docker 8.96 seconds) enables concurrent agent deployment (WEBSERV, Table 2).
5. Accessibility tree navigation enables agents to recognize repeated actions and maintain structured interaction state (WEBSERV, Section 3.2.1).

Real Limitations Acknowledged Only in Asides

Gender/age bias in participant pools (skew toward 35+ males in tech). Single-platform bias (Taobao, Lazada, Amazon-specific findings may not generalize). Language bias (German dominance in OptAgent, with "Others" category showing only 0.49% improvement, Table 2; German shows 3.59%). Incognito-mode exclusion. Enterprise user exclusion. Privacy-conscious user self-selection bias.

These gaps are not speculative. Each emerges from the intersection of what the twenty-two papers collectively measure and what they collectively omit.

• Real-Time Cross-Platform Price Optimization: No system evaluates dynamic pricing. AgenticShop documents that 70% of price failures are out-of-budget recommendations (Table 6). OptAgent optimizes static queries.
• Trust and Reputation Modeling: No paper models customer trust in agent recommendations. Yang et al. (LLM Agent Meets Agentic AI) documents that humans express concerns about bias and manipulation that agents never report.
• Seasonal and Temporal Shopping Patterns: All evaluations use static datasets.
• Multi-Session User Preference Learning: All systems treat sessions as independent.
• Fraud Detection and Adversarial Scenarios: No benchmark includes fraudulent sellers or counterfeit products. The Agentic Web Survey catalogs 16 threat types but shopping-specific fraud detection is absent.
• Multicultural and Multilingual Shopping: OptAgent shows language-dependent performance (German 3.59% vs Others 0.49%, Table 2).
• Accessibility and Inclusive Design: Zero papers evaluate agents for users with disabilities.
• User Preference Elicitation Mechanisms: No paper addresses how agents should ask clarifying questions. TRACER uses decision trees but does not compare to other strategies.

Four stakeholder groups face distinct strategic imperatives. The research base supports specific recommendations for each. Evidence strength labels follow the companion translations: claims marked as projections or inferences are identified.

The twenty-two papers tell one story: agentic shopping works at population scale, fails at individual scale, and breaks predictably under complexity. The path forward is not larger models: it is better systems. Better retrieval. Better decomposition. Better safety infrastructure. Better personalization architecture.

The governance opportunity is clear: the Autonomy Levels framework provides structure for calibrating agent autonomy as a deliberate design decision. Current production systems operate at L3–L4 (partial autonomy with human approval); full autonomy (L5) remains risky for financial transactions until the approximately 40-point safety gap closes.