PaperlayerView exact revision on arXiv

Source-linked AI summary

Developer Experiences with Cursor AI-native IDE

Haoran Xue, Gias Uddin, Song Wang

arXiv:2610.20618v1cs.SE

TL;DR

Prior research offers limited longitudinal, community-driven evidence on how developers experience AI IDEs as everyday work environments. This study analyzes more than two years of posts from Cursor’s public forum. It finds that workflow-critical mechanisms are central to both perceived value and frustration, while friction shifts over time from access barriers toward pricing and robustness.

  • Problem

    Prior evidence based on controlled tasks, small interview samples, and short observations offers limited visibility into longitudinal, community-driven AI IDE use and workflow drift.

    Source: §1 · p. 1

  • Method

    The study analyzes 14,714 posts from Cursor’s official public forum using a taxonomy, sentiment, usage-posture, and interaction signals for longitudinal analysis.

    Source: §1 · p. 2, §8 · p. 19

  • Results

    Workflow-critical mechanisms, including context/indexing, model routing, and agent workflows, are both primary sources of value and dominant failure surfaces, while attention shifts from access barriers to capability limits and later pricing/value and robustness.

    Source: • RQ3: Where Does Cursor Break Down? · p. 2, §8 · p. 19

  • Takeaways & Limitations

    The findings motivate design priorities including operational readiness, transparent and controllable context, intelligible routing, bounded agent autonomy with recovery paths, and trustworthy usage accounting.

    Source: §8 · p. 19

    The findings reflect publicly expressed experiences in one product community and may not generalize to silent users, private teams, or enterprise settings.

    Source: §7 · p. 18

Abstract

from arXiv · show

AI-powered integrated development environments (IDEs) are increasingly positioned as practical tools for software development, integrating AI assistance into everyday activities such as editing, refactoring, and debugging. However, how developers evaluate AI IDEs as working environments, what frictions shape their adoption, and which aspects of the experience matter most in everyday use remain underexplored. Prior evidence on adoption and use is mainly controlled tasks, small interview samples, and short observation windows, which offer limited visibility into longitudinal, community-driven experience and the workflow drift that emerges in everyday practice. To bridge this gap, we conduct a large-scale longitudinal study of Cursor, a prominent AI-native IDE, through its official public forum, where users report bugs, request features, and discuss everyday use. Our analysis covers 14,714 forum posts collected over more than two years. We find that discussion is dominated by workflow-critical mechanisms (71%) such as context indexing, model routing, and agent workflows, which are also the main sources of both perceived value and frustration. Negative experiences center on usability, pricing, and dependability, including installation and setup, usage limits and consumption, crashes, and latency. Over time, the focus of friction shifts from early access and setup barriers toward pricing, robustness, and sustained workflow reliability.

1 Introduction

Prior studies offer limited visibility into longitudinal, community-driven use of AI IDEs as evolving work environments. This study analyzes Cursor’s public forum to characterize developers’ experiences and how they change over time.

1 Introduction

  • Prior work often uses controlled tasks, small interviews, or short observations, limiting visibility into cumulative workflow effects and tooling drift over months or years.Earlier studies also tend to examine assistants as point tools rather than IDEs as evolving work environments.Source: §1 · p. 1
  • The analysis uses open coding to build a taxonomy of nine categories and 34 subcategories, then examines attention and positive and negative discourse over time.Attention is quantified using views, likes, and comments; sentiment prevalence and interaction patterns are traced over time.Source: • RQ3: Where Does Cursor Break Down? · p. 2
  • Functionality and usability dominate discussion and are also major sources of negative experiences, alongside access, pricing, and credits.Complaints include usage limits and unchecked consumption; reported negativity is 88.1% for performance and 68.7% for dependability.Source: • RQ3: Where Does Cursor Break Down? · p. 2
  • The study reports that a growing share of forum posts describe Cursor as the primary IDE, while negative discussion shifts from usability and functionality toward pricing and dependability.The primary-IDE share rises from 13% in 2023Q3 to 26.6% in 2025Q4.Source: • RQ3: Where Does Cursor Break Down? · p. 2
  • The study analyzes 14,714 Cursor forum discussions collected over more than two years to provide a longitudinal, community-grounded account of developers’ experience with an AI-first IDE.The forum is a product-specific venue where users report bugs, request features, discuss usage, and evaluate the tool in everyday practice.Source: • RQ3: Where Does Cursor Break Down? · p. 2, §1 · p. 2
  • The stated contributions include design lessons for AI-first IDEs spanning readiness and trust, legibility and control, and workflow.The supplied passage ends before the full list of lesson areas.Source: • RQ3: Where Does Cursor Break Down? · p. 2

2 Background and Related Work

Prior work considers how IDE design and AI assistance affect practical development, but findings on user perceptions come from varied methods and settings. This background frames the study’s focus on AI-native IDEs as work environments and on adoption in practice.

2.1 IDE Design

  • AI-native IDEs extend traditional environments with richer context handling, tighter workflow support, and more agentic interaction models.Prior systems research emphasizes transparency, control, correctness, and workflow fit as conditions for practical value.Source: §2.1 · p. 3

2.2 Studies on User Perceptions of AI Coding Assistants

  • Prior user-perception research identifies recurring barriers including reliability, incomplete outputs, weak project context, confusing onboarding, resource overhead, and pricing.Marketplace reviews and surveys also report concerns about inaccuracy, trust, contextual understanding, and workflow friction.Source: §2.2 · p. 3
  • Adoption and performance findings vary with usefulness, ease of use, trust, risk, task, tool, expertise, and validation effort.Positive attitudes can coexist with limited performance gains when understanding, validating, and repairing generated code takes substantial effort.Source: §2.2 · p. 3

3 Study Design

The study combines forum-post collection, hybrid topic labeling, sentiment analysis, interaction measures, and usage-posture annotation to examine Cursor use at scale. The dataset spans more than two years and includes 14,714 posts.

3.1 Data Collection

  • 14,714 forum posts were collected across 813 days, from August 22, 2023, to November 13, 2025.The dataset also includes 51,965 comments, 11.18 million views, and 51,518 likes.Source: §3.1 · p. 3

3.2 Data Labeling

  • Researchers manually annotated 1,000 posts to develop a guide, then used GPT-5 mini in a human-in-the-loop workflow to label the remaining posts and refine a two-level taxonomy.The process retained open-coding interpretability while scaling annotation across the corpus.Source: §3.2 · pp. 4–5

3.3.1 Sentiment Analysis.

  • Sentiment labels distinguish positive, neutral, and negative posts; a manual check of 1,000 labels found 81% overall agreement.The labels capture approval, non-evaluative description or inquiry, and dissatisfaction, respectively.Source: §3.3.1 · p. 5

3.3.2 Attention and Engagement Measures.

  • Popularity, agreement, and engagement are measured as views, likes, and comments, respectively, and quarterly means summarize average interaction intensity rather than typical thread behavior.The analysis reports quarterly post counts to contextualize periods with low volume.Source: §3.3.2 · p. 5

3.3.3 Usage posture labeling.

  • Usage posture is classified from post-level cues into primary_ide, non_primary_ide, trial_blocked, or unclear; LLM-based labeling achieves 86% agreement against human validation.Auditability fields include adoption preconditions, tradeoffs, IDE change events, and brief evidence snippets quoted from the thread.Source: §3.3.3 · pp. 5–6

4.1 RQ1. What Topics Do Users Discuss When Using Cursor?

The taxonomy covers nine high-level categories and 34 subcategories, with discussion concentrated on workflow-critical capabilities and the frictions around using them. The most discussed capabilities are also prominent sources of both praise and pain.

4.1.1 Functionality and Capabilities.

  • Functionality & Capabilities topics cover repository context, model selection, agent editing across files, editor behavior, and external-tool integration.These mechanisms shape how Cursor supports real coding tasks.Source: §4.1.1 · p. 6

4.1.2 Usability and Access.

  • Usability & Access covers installation, login, configuration, interface controls, remote setups, integrations, and enterprise governance.These are operational-readiness conditions that can create friction before users reach core capabilities.Source: §4.1.2 · p. 6

4.1.3 Pricing and Plan.

  • Pricing & Plan discussions address quotas and usage limits, tier value, billing transparency, and purchasing logistics.These include requests for clarity around usage accounting and plan semantics.Source: §4.1.3 · p. 6

4.1.4 Dependability and Stability.

  • Dependability & Stability includes agent hangs and loops, regressions, stuck background tasks, crashes, frozen interfaces, and broken configurations.Performance & Resource Usage separately captures latency, stalls, network reliability, and resource consumption.Source: §4.1.4 · p. 6, §4.1.6 · p. 7

4.1.5 General Experience.

  • The taxonomy also includes holistic evaluations and community discussion, comparisons with other tools, privacy and execution-safety concerns, and forward-looking feature or documentation proposals.Future Suggestions posts are typically constructive; privacy discussions include data handling, compliance, and agent command controls.Source: §4.1.5 · p. 6, §4.1.7 · p. 8, §4.1.8 · p. 8

4.1 RQ1. What Topics Do Users Discuss When Using Cursor?

  • Context & Indexing, Models & Routing, Agents & Workflows, and UI & Layout dominate both users’ liked and disliked topic lists, making the same workflow-critical features sources of value and friction.When they work, users credit productivity and workflow fit; failures can break workflows through wrong context or confusing interactions.Source: §4.1.9 · p. 8

4.2 RQ2. What Do Developers Favor About Cursor?

The authors examine positivity across product topics, forum sections, interaction signals, and usage postures. Positive feedback is unevenly distributed, while negative experiences more often prompt discussion and troubleshooting.

4.2.2 Results.

  • 46.4% of Feature Requests posts are positive, and they account for 61% of all positive posts; Bug Reports are almost entirely negative.Positive evaluations are often embedded in proposals to extend or refine features that already work well.Source: §4.2.2 · pp. 8, 10
  • Functionality & Capabilities and Usability & Access are prominent on both sides of sentiment: they contribute about 73% of positive posts, while the category-level counts include 1,338 positive versus 2,576 negative for Functionality and 566 versus 1,922 for Usability.Performance & Resource Usage and Privacy & Security are overwhelmingly negative, at 1.6% and 4.3% positive, respectively.Source: §4.2.2 · pp. 9–10
  • Positive posts shift from early high per-post visibility to later higher volume but lower per-post visibility, while negative posts draw comparable or greater comment engagement.Popularity and agreement favor positive posts overall, whereas replies are closer to parity and sometimes favor negative posts.Source: §4.2.2 · p. 11
  • 33.0% of primary-IDE posts are positive, compared with 13.2% of non-primary-tool posts and 2.3% of trial-blocked posts.Positive experience is concentrated among users who have integrated Cursor into their daily development loop.Source: §4.2.2 · pp. 11–12
  • The share of posts describing Cursor as a primary IDE rose from 13.0% in 2023Q3 to 26.6% in 2025Q4, while non-primary use remained at 9–13%.Trial-blocked posts also increased from 1.0% to 5.4%.Source: §4.2.2 · pp. 12–13

4.3 RQ3. Where Does Cursor Break Down?

The analysis examines negative experience by post volume and per-post interaction, finding that breakdowns cluster in a few high-volume or high-attention areas. Negative attention shifts over time from access and usability toward capability, pricing, and dependability concerns.

4.3.2 Results.

  • 6,570 of 14,714 posts (44.7%) are negative, spanning 34 subcategories, and negative experience is concentrated unevenly across subcategories.The analysis distinguishes how often problems occur from their per-post interaction intensity.Source: §4.3.2 · p. 13
  • Install, Updates, Packaging & System Requirements and API Keys & Connectivity combine high negative-post volume with about 1.2k views per negative post.Figure 9 compares mean views, likes, and comments; bubble size represents post volume.Source: §4.3.2 · p. 13, Figure 9 · p. 14
  • Pricing & Plan has 1,395 posts, with 8% positive and 45% negative posts; users complain about usage limits, consumption, and unclear rate-limit and price visibility.Complaints include confusion and perceived unfairness around usage accounting and plan semantics.Source: §4.3.2 · pp. 13–14
  • Crashes and state corruption draw 98% negativity, while latency and stalls draw 94% negativity and high resource usage has 87% negativity.High resource usage has 5% positivity; crashes and state corruption have 0% positivity, as do latency and stalls.Source: §4.3.2 · p. 14
  • Negative attention initially centers on Usability & Access, then shifts toward Functionality & Capabilities near the 2025Q1 attention peak.Early attention is consistent with access, setup, and first-order usability issues; capability complaints later concern core features and workflow-critical tasks.Source: §4.3.2 · pp. 14–15
  • By 2025Q3 Pricing & Plan gains a large share of negative views despite a smaller attention baseline, while 2025Q4 attention shifts toward Dependability & Stability and General Experience.Negative views decline sharply in recent quarters, and the drivers of negative volume differ from those of negative attention.Source: §4.3.2 · p. 15

5 Discussion

Comparison discussions show different priorities depending on the reference tool: continuity with established environments and stronger AI support with AI competitors.

5 Discussion

  • VS Code accounts for 69.7% of comparison posts, which emphasize workflow continuity, editor experience, extension compatibility, and terminal or remote-environment use.The comparison set also includes traditional IDEs (5.5%), AI coding assistants (21.9%), and other AI-powered IDEs (2.9%).Source: §5 · p. 16
  • Comparisons with AI coding assistants and AI-first IDEs focus on model selection, routing, pricing, API flexibility, and the quality of the AI-mediated workflow.Across comparison types, the reported priorities distinguish low-friction adoption from stronger, more predictable, and better-valued AI support.Source: §5 · p. 16

6 Lessons Learned

The lessons learned group findings into three themes: readiness and trust infrastructure, legibility and control of AI mediation, and sustaining end-to-end workflow.

6 Lessons Learned

  • The paper synthesizes its findings into six takeaway messages grouped under three themes.The themes address readiness and trust infrastructure, AI mediation, and end-to-end workflow.Source: §6 · p. 16, Table 4 · p. 18

6.1 Readiness and Trust Infrastructure

  • Installation, login, provider connectivity, plan recognition, and recoverability can prevent developers from evaluating core AI features when basic readiness fails.The authors recommend robust first-run flows, actionable errors, clear fallbacks, and state recovery as core product requirements.Source: §6.1 · p. 16
  • Opaque usage accounting can undermine confidence in predictability and fairness, so pricing is treated as part of the user-experience contract.The authors call for legible costs, verifiable accounting, and transparent gating at the point of interaction.Source: §6.1 · p. 16

6.2 Legibility and Control of AI Mediation

  • Because context and indexing are both valued and criticized, the authors recommend making context construction visible, steerable, and verifiable.Users should be able to inspect indexed sources and constrain context through rules, project boundaries, and exclusions.Source: §6.2 · p. 17
  • Model routing and negative performance sentiment make transparent selection policies and predictable responsiveness central design requirements.The authors recommend controls for prioritizing speed, quality, or cost, and graceful degradation to avoid costly context switching.Source: §6.2 · pp. 17–18

6.3 Sustaining End-to-End Workflow

  • Agentic workflows need inspectable plans, staged and revertible edits, and clear pause and recovery mechanisms to limit downstream repair costs.The authors connect reliability and recovery, rather than autonomy alone, to incorporating agents into everyday development.Source: §6.3 · p. 18
  • Ecosystem parity and low switching costs matter because comparisons with VS Code emphasize compatibility and continuity across everyday development tasks.The authors identify extension compatibility, settings migration, terminal and remote-development stability, and predictable integrations as product requirements.Source: §6.3 · p. 18

7 Threats To Validity

The study’s findings are bounded by its public-forum data, annotation process, engagement proxies, and focus on one AI-native IDE.

7 Threats To Validity

  • The public-forum results reflect publicly expressed Cursor user experiences and may not generalize to silent users, private teams, or enterprise settings.The study covers more than two years but represents one product community.Source: §7 · pp. 18–19
  • Some ambiguous or mixed-intent posts may be misclassified; sentiment labeling reached 81% correctness and usage-posture labeling 86% agreement against human validation.Engagement signals and quarterly aggregation are imperfect proxies for importance, and results from one AI-native IDE should be generalized with care.Source: §7 · p. 19

8 Conclusion

The study analyzes more than two years of Cursor forum discussion to characterize developers’ experiences. Workflow-critical mechanisms are both primary sources of value and dominant sources of frustration, while discussion shifts over time from access barriers toward pricing, value, and robustness.

8 Conclusion

  • The study analyzes 14,714 Cursor forum posts collected over more than two years using a two-level taxonomy of nine categories and 34 subcategories, alongside sentiment, usage-posture, and interaction signals.Source: §8 · p. 19
  • Workflow-critical mechanisms are both primary sources of value and the dominant failure surface, with attention shifting from access barriers toward capability limits, pricing/value, and robustness.The findings motivate priorities including operational readiness, transparent and controllable context construction, intelligible routing, bounded agent autonomy with recovery paths, and trustworthy pricing and usage accounting.Source: §8 · p. 19
  • The authors provide the dataset and source code online.Source: §8 · p. 19
Loading 2610.20618v1…