Today we learned Cursor’s “customer support” is actually an LLM pretending to be human, misleading many users. As a substitute of proudly owning their mistake, Cursor doubled down, mendacity and deleting the dialogue from their subreddit to suppress this- a tactic they’ve used earlier than.
Since I first heard about it, I’ve extensively examined Cursor, repeatedly flagging stability and safety points, solely to get ghosted every time. I gave them the advantage of the doubt (they’re very fashionable, so my message may’ve gotten buried). I figured they’d take heed to neighborhood suggestions to repair their shit. However their newest cover-up echoes the sketchiness of Devin AI, which, coincidentally, we wrote about this time last year. One thing about April brings out the three S’s: Spike-ball (GOATed sport, invite me please), Sundresses, and Scammers.
If that’s what they need, who am I to say no?
Cursor’s rise has been meteoric, pushed by Andrej Karpathy’s endorsement and a fear-of-missing-out pitch — “Your rivals use Cursor; don’t fall behind.” They cruised previous $100M ARR in document time-
However enterprise software program isn’t a weekend hackathon. Legacy codebases, intricate dependencies, safety, and compliance guidelines create a particular sort of monster. And on this context, Cursor isn’t simply inconvenient- it’s doubtlessly harmful.
This text explores why Cursor is a poor match — and dangerous — for critical enterprise coding. For no matter my opinion is value, I’d strongly urge (and have in all my related conversations) that professionals keep away from Cursor for different alternate options b/c of the numerous points we are going to talk about. Till they make some critical adjustments, Cursor must be prevented, even when it’s given away totally free (you’re higher off shopping for a greater software). It does do some issues properly (some very properly, even), however there are higher instruments on the market, and any person ought to contemplate migrating to one among them.
Hello since i do know you’ll by no means reply to this or hear this.
We spent nearly 2 months combating with you guys about fundamental questions any B2B SaaS ought to be capable of reply us. Issues resembling invoicing, contracts, and safety insurance policies. This was for a low 6 determine MRR deal.
When your gross sales rep responds “I don’t know” or “I might want to get again to you” for weeks about fundamental questions it left us with a large disappointment. Please do higher, nonetheless now we have moved to Copilot.
–When a customer complaint starts with you’ll never hear or respond to this, you’re crossing even Google’s levels of sloth. For a startup, that is unacceptable and the whole Cursor workforce must be made to look at Arjun Kapoor motion pictures on loop.
An vital disclaimer earlier than we begin-
I like AI Code
This isn’t a campaign towards AI-generated code assistants. I depend on AI Code extensively. At the moment, my work makes use of this in 3 capacities (so as of significance, most vital first)-
- Augment– I’ve a easy, however efficient workflow with it. Activate chat mode (I don’t like their agent), ask it questions in regards to the code base (“Hey I need to do that, what points matter probably the most/what would I want to alter”), and make options based mostly on the approaches for fixing (“this strategy appears needlessly sophisticated/will should be mounted, can we alter the return to this…”) . Repeat as many instances as wanted. Increase’s superior understanding of the code base implies that each one-shot era for easy adjustments and three–5 iterations of asking + reviewing earlier than accepting adjustments, Increase simply does a greater job. You’ll be able to really feel a really huge distinction in high quality for his or her AI Strategies between Increase and everybody else. Increase additionally has glorious debugging capabilities usually. Their 30 USD/month plan is an absolute steal and I’d advocate it to everybody (this isn’t sponsored, and I get no monetary compensation from them- I similar to the product). Nevertheless, if the Increase workforce needs to slide me some share factors in fairness in an unmarked envelope, I’ll settle for, purely for analysis functions.
- o1 with Deep Analysis- This helps lots with my analysis. If I have to prototype an strategy based mostly on papers/some concepts I’ve, I simply have o1 do one shot generations with DR. DR permits you to add PDFs and it’s fairly good at following directions, so I’ve been pleased with the outcomes, even after I have to debug. I’ve tried Gemini and Perplexity Deep Analysis, however they’re each very mediocre. Iqidis has a phenomenal Deep Research tool (nudge nudge) which is FREE!, but we specialized that one to focus on law, so it’s not as good for AI coding (try it out here though, even for non law queries, you might be pleasantly surprised). I’m excited to check out o3 and see how properly it does as compared.
- Claude, o1, and a pair of.5 Professional for random debugging -Typically Increase isn’t good for single file debugging so I take advantage of these fashions (there isn’t actually a rating, every does properly generally and no sample as of now).
All 3 save me numerous time (particularly the primary 2), and I’ll write my information on utilizing them finally.
This text argues that whereas the AI coding assistant Cursor reveals potential and velocity, its present design and AI behaviors make it unsuitable and dangerous for enterprise software program growth on account of elementary conflicts with enterprise necessities. Specifically:
Enterprise Context is Completely different:
Not like demos or startups, enterprise growth includes huge legacy codebases, excessive stakes (monetary, safety, compliance), advanced established workflows (monorepos, CI/CD, opinions), and numerous groups. Stability, safety, and maintainability are paramount.
Cursor’s Core Design Creates Main Issues in Enterprise:
- Unmanageable Workflow & Output: Its automated multi-file adjustments, such because the vaunted “Agent Mode” generates large, messy, hard-to-review Pull Requests, crippling the code overview course of and resulting in frequent, painful merge conflicts for groups.
- Unreliable AI Habits: The AI usually produces damaged logic (“code on LSD”), hallucinates syntax/APIs, makes unintended adjustments outdoors the requested scope (generally harmful, like resetting databases regardless of guidelines), and degrades architectural high quality over time by introducing inconsistencies and tech debt.
- Foundational Blockers: Important points embrace main safety dangers (sending proprietary code externally, failing to reliably ignore delicate information like .env), compliance failures, and poor efficiency/instability (sluggishness, crashes, excessive useful resource utilization) on giant, real-world enterprise codebases.
- Poor Buyer Assist: Reviews of unreliable AI help brokers offering misinformation, gradual/non-existent human help for paying customers, and lack of enterprise-grade responsiveness erode belief.
All in all, a excessive perceived value, restrictive limits on options, and the intensive want for human oversight/rework make the ROI doubtful in comparison with alternate options (like Copilot or VS Code plugins utilizing letting customers use their very own API keys).
On a extra conceptual/deeper stage, Cursor has 3 important points that should be addressed-
Incompetence of constructing Information-Intensive AI
Wanting on the outputs and issues that Cursor has, particularly for giant, advanced code bases, it appears that evidently the Cursor workforce is in over their heads relating to constructing AI for classy information work. In such contexts, you possibly can’t simply throw a bunch of stuff into an listed database, name an AgentSmith workflow, and name it a day.
Constructing refined reasoning/analysis frameworks that permit your AI to course of, retrieve, and self-correct on suggestions could be very troublesome work and it requires numerous analysis to successfully decide the few key choices that drive outsized returns.
It’s not a coincidence that their issues are continually associated to the tougher offerings- retrieval and evaluation on giant code bases, brokers, balancing particular rule following with creativity, and so on. There’s solely up to now that counting on model-level intelligence can take you for KI-AI, and the workforce Cursor doesn’t have the flexibility to interrupt by the wall. That’s why they do numerous issues, however fail to do them properly.
Evaluate them to Increase and even Claude Code. Each have a lot stronger groups and experience, which is why you’re feeling a qualitative distinction within the high quality of their solutions for advanced queries.
The “Fool-Proof” Downside: Unsafe for Numerous Groups:
This isn’t a “flaw” with the product, however nonetheless an vital consideration when contemplating enterprise use circumstances. Instruments used throughout enterprises should be “idiot-proof”- protected even when not operated completely by each single person. This reduces the oversight we’d like, permitting to to actually unlock positive factors for decrease talent workers with out burdening the upper talent ones with overview.
Cursor’s design isn’t strong sufficient for dependable use throughout groups with various talent ranges and diligence, making it susceptible to misuse.
- Inadequate Management & Transparency: It affords an “phantasm of management” with options like “Guidelines” which might be reportedly ignored. Key actions like context choice and multi-step edits usually occur opaquely, with out enough previews or necessary person affirmation factors for important adjustments.
- Encourages Dangerous Over-Reliance: The convenience of triggering broad, automated adjustments with out clear, enforced checkpoints makes it too simple for any workforce member (particularly these beneath strain or much less skilled) to inadvertently introduce errors or safety lapses which might be laborious to catch.
- Overloads Human Oversight: The mix of huge, messy, doubtlessly flawed AI output locations an unrealistic overview burden on workforce members, undermining the effectiveness of code opinions as a security web when the producing software itself lacks enough safeguards.
No testing.
Fairly simple- they don’t have testing setup. Use a few of that funding cash to implement correct chaos engineering and QA. Ship me a message in order for you entry to good high quality testers, reviewers, or focus groups- they’ll do significantly better than no matter you’ve got occurring proper now.
Finally, Cursor is a superb instance of workforce (one has to respect the workforce for constructing such a viral product) making an attempt to unravel an issue that’s an excessive amount of for them. Its fame amongst influencers, startups, and vibe coders is not any coincidence, because it offers them what they need- simple wins w/o fear for the longer term. Nevertheless, it’s not a critical individuals, and shouldn’t be used the place critical persons are wanted. Whereas it may be superb in the best palms, it has too many limitations within the enterprise setting to make it suggestion.
The hype round AI coding is actual, however so are the dangers. Let’s reduce by the noise and look at why Cursor, in its present type, is likely to be extra of a legal responsibility than an asset when the stakes are excessive.
I put numerous work into writing this article. To take action, I depend on you for help. If a couple of extra individuals select to change into paid subscribers, the Chocolate Milk Cult can proceed to supply high-quality and accessible training and alternatives to anybody who wants it. Should you assume this mission is value contributing to, please contemplate a premium subscription. You are able to do so for lower than the price of a Netflix Subscription (pay what you want here).
I present varied consulting and advisory companies. Should you‘d prefer to discover how we will work collectively, reach out to me through any of my socials over here or reply to this electronic mail.
Earlier than we dive into Cursor’s particular challenges, it’s essential to grasp why enterprise software program growth operates on a unique airplane than private tasks, small startups, or remoted code demos. It’s not merely a matter of “extra traces of code.” This may present essential context for what does and doesn’t make good coding assistants for skilled software program growth settings.
Not like greenfield tasks, the place you’re constructing from scratch, enterprise builders usually work inside large, interconnected codebases — usually a long time previous and wayy too many traces lengthy. These programs are riddled with technical debt, undocumented corners, and sophisticated dependencies that may make even seemingly easy adjustments a dangerous enterprise. Any software touching this code must tread rigorously and perceive the prevailing context, not simply generate new, remoted snippets.
Second, the stakes are astronomically increased. A bug in a private undertaking is likely to be a minor annoyance. In an enterprise system, it could actually result in monetary losses, reputational harm, regulatory fines, and even safety breaches affecting thousands and thousands of customers. This implies there’s zero tolerance for instruments that introduce instability or compromise knowledge integrity. On this dynamic, the price of failure is given a a lot increased weightage than productiveness will increase (consider enterprises are by their nature, risk-averse)
Third, safety and compliance are paramount. Enterprise code isn’t simply mental property; it usually handles delicate knowledge and should adhere to stringent rules (SOC2, GDPR, HIPAA, and extra). Information sovereignty, secrets and techniques administration, and strong entry controls are non-negotiable. Instruments that transmit code to exterior servers with out clear safety ensures or violate compliance requirements are lifeless on arrival.
Sure and it irritates the hell out of me. Cursor help is rubbish, however points with billing and different issues are a lot worse.
The workforce I work with it took almost 3 months to get fundamental questions answered appropriately when it got here to a gross sales contract. They by no means gave our Sec workforce acceptable solutions round privateness and safety.
Fourth, enterprise growth depends on entrenched workflows and sophisticated toolchains. Code doesn’t exist in a vacuum. It flows by refined pipelines involving model management (Git), construct programs, steady integration and steady deployment (CI/CD), static evaluation, automated testing, and extra. A brand new software should combine seamlessly with these present programs, not disrupt them or power builders to undertake solely new processes. This consists of issues like monorepo help, IDE plugin ecosystems, and compatibility with established coding requirements.
Lastly, there’s the human ingredient. Enterprise groups are composed of builders with various talent ranges, area experience, and familiarity with completely different elements of the codebase. Builders will stop halfway by, leaving half-completed chains of their wake. All of those unfastened ends create numerous points for Code Assistants, which get very confused and can usually miss vital context, or attend to the unsuitable components within the code.
In brief, enterprise software program growth calls for stability, safety, maintainability, and rigorous course of management. It’s an setting the place even small errors can have monumental penalties. This creates a excessive bar for any software in search of to enhance developer productiveness — a bar many AI coding assistants, together with Cursor, battle to clear.
“Cursor particularly was one of many worst; the very first time I allowed it to have a look at my codebase, it hallucinated a lacking brace (my code parsed fantastic), “helpfully” inserted it, after which proceeded to interrupt every part. How am I alleged to belief and work with such a software? To me, it looks like the equal of lobbing a stay hand grenade into your codebase.
… I really feel the hallucinations may be off the charts — inventing APIs, perform names, whole libraries, and even whole programming languages from time to time. The AI is very happy to ship any sort of data you need, irrespective of how unsuitable it’s.
AI isn’t a software, it’s a tiny Kafkaesque paperwork inside your codebase. Does it work right now? Sure! Why does it work? Who can say! Will it work tomorrow? Fingers crossed!”
Now that we’ve established the demanding nature of enterprise growth, let’s look at how Cursor’s particular options and AI behaviors usually fail to fulfill these necessities.
A. The “AI Hand Grenade”: Producing Unreviewable Pull Requests, Messy Diffs, and Merge Nightmares
Enterprise growth isn’t a solo endeavor. A number of builders, usually on completely different groups, work concurrently inside the identical giant codebase. This makes cautious coordination and clear integration paramount. Right here, Cursor’s tendency to generate large, automated adjustments turns into notably problematic.
One in all its fundamental promoting factors is its means to carry out multi-file edits mechanically particularly for those who set off its “Agent Mode.” Give it a activity, and it could actually contact quite a few elements of the codebase. The rapid drawback? This usually ends in sprawling, monolithic pull requests (PRs). Builders have reported the AI altering greater than it ought to (“I end up with changes in random files I never intended to touch”). When builders aren’t cautious, this could result in PRs so giant they change into virtually unreviewable-
These PRs are every a thousand traces lengthy. If anybody hasn’t skilled reviewing giant quantities of AI-generated code earlier than, I’ll inform you it’s like studying code written by a schizophrenic. It takes numerous effort and time to make sense of such code and I’d moderately not be reviewing coworkers’ AI-generated slop and being the one one stopping the codebase from spiraling into being utterly unusable
This isn’t simply a person reviewer’s headache; it’s a collaboration nightmare. When a number of builders on a workforce are utilizing Cursor to generate giant, overlapping adjustments concurrently, the chance of painful merge conflicts skyrockets. Not like smaller, targeted human commits, these broad AI-driven adjustments improve the floor space for collision dramatically. Resolving these conflicts isn’t simply time-consuming; it’s fraught with danger, as builders battle to manually combine advanced, AI-generated logic from completely different branches.
Moreover, the diffs themselves are sometimes messy, exacerbating merge difficulties. This pointless churn makes automated merges much less more likely to succeed and guide merges considerably tougher and extra error-prone.
The mixed impact is devastating for workforce velocity. As a substitute of accelerating growth, Cursor can grind it to a halt with unreviewable PRs that block progress (“kills momentum”) and frequent, advanced merge conflicts that devour developer time and introduce errors. In an enterprise setting that values clean integration and predictable progress, this can be a important step backward from the cautious, incremental adjustments usually favored.
B. Logic on LSD: Damaged Management Flows and Hallucinated Code
AI fashions, even highly effective ones, lack a real understanding of program logic. Cursor is not any exception. Builders persistently report situations the place the AI introduces damaged or nonsensical management flows. We’ve lined a number of examples of this by now.
This erratic habits stems from the AI’s lack of ability to know the deeper context and architectural constraints. A Scala developer likened the experience to working with “a developer who is on LSD,” citing examples of Cursor attempting to rewrite libraries or override functions randomly. The AI may generate code that compiles however accommodates refined logical flaws, bypasses present safeguards, or fails to propagate adjustments appropriately throughout all needed utilization websites.
In enterprise programs, refined logic bugs like these change into ticking time bombs. In our expertise, Cursor has often bypassed whole authentication flows — or, worse, silently duplicated them, scattering redundant logic throughout completely different elements of the codebase. Such points are dangerously simple to miss, particularly when builders aren’t persistently reviewing or interacting with the whole system.
C. Collateral Injury Inc.: Unintended Aspect Results Throughout the Codebase
Cursor’s means to mechanically index and modify code throughout a whole repository is a double-edged sword. Whereas supposed to allow highly effective refactoring, it continuously results in unintended unintended effects, particularly in giant tasks or advanced monorepos. The duplication of logic talked about earlier is a distinguished and harmful instance of this.
Extra alarming are circumstances the place the AI ignores express constraints. One developer reported Cursor repeatedly resetting their database regardless of “Guidelines” set as much as stop it-
Even the much less dramatic unintended effects, just like the AI “going off script” to refactor unrelated code it deems suboptimal, violate the precept of least shock. Builders are compelled to hunt by doubtlessly quite a few information to evaluate collateral harm after each AI operation.
This wipes out a lot of the productiveness positive factors that utilizing AI to generate the solutions would offer to start with.
D. The Black Field Downside: Safety, Compliance, and Efficiency Failures
Compounding the very evident situation of Cursor sending Information to 3rd get together severs (and LLMs) was the alarming discovery that Cursor doesn’t all the time respect ignore information (.gitignore, .cursorignore), sending delicate knowledge like .env information containing API keys and secrets and techniques to exterior servers-
This isn’t a minor situation that may be neglected, and even when it’s already been patched, the truth that such a evident situation handed by the workforce isn’t one thing that instills numerous confidence of their workforce/product.
Add to this an absence of transparency surrounding knowledge dealing with (“by no means acquired acceptable solutions round privateness and safety”), which makes it unattainable for enterprise safety groups to vet and approve the software. Hopefully, the issues are beginning to settle in.
Lastly, even when safety had been good, efficiency points plague Cursor, particularly on giant enterprise codebases. Reviews abound of the editor turning into “extremely sluggish,” freezing, crashing continuously (“greater than 5 instances a day”), or exhibiting important typing lag. The codebase indexing required for AI context can take terribly lengthy and even loop indefinitely, forcing customers to disable it and lose the very function Cursor emphasizes.
Excessive CPU, reminiscence, and battery drain are additionally widespread complaints. If the software is unstable or painfully gradual on the precise tasks builders work on, it doesn’t matter how good the AI is — it’s unusable. These foundational points characterize unacceptable dangers and sensible blockers for critical enterprise adoption.
Did Cursor present promise? Sure. Has a lot of their habits and up to date updates have proven an entire misunderstanding of the software program growth course of, a regarding lack of ethics, and a misunderstanding of the right way to construct AI options for information work? Additionally, sure. To me, the chance of investing in and utilizing it’s merely not value it.
In some ways, Cursor embodies the present era of Generative AI startups — constructed by founders who lack real experience in each AI and their claimed domains, propped up by the TikTokification of AI analysis, the place two-minute explainers and shallow hot-takes from on-line hustlers have change into the default supply of knowledge. Penning this, I can’t assist however recall a VR combat-sports startup I encountered some time in the past. They marketed their product as an immersive coaching software by providing a VR headset that supposedly enhanced shadowboxing realism by taking part in prerecorded movies of fighters throwing pictures.
In actuality, it was a clunky toy: failing to seize the fluidity, dynamism, and adrenaline spike of precise sparring whereas additionally limiting motion greater than plain previous shadowboxing would. Designed by fits who‘d by no means taken a punch and praised enthusiastically by individuals who threw imply left hooks and clean step-through knees on their keyboards. Someplace down the road, the old style folks who really stepped into the cage had been utterly neglected.
Or who is aware of, possibly I’m unsuitable about the entire thing and Cursor is definitely good. In any case, it’s not as if the VC-Tech-Influencer Circle Jerk has ever overestimated a product and misled the individuals before-
Let me know what you assume.
Thanks for being right here and I hope you’ve got a beautiful day,
Please vibe your means into becoming a member of my cult,
Dev ❤
I put numerous effort into creating work that’s informative, helpful, and impartial from undue affect. If you’d like to support my writing, please consider becoming a paid subscriber to this newsletter. Doing so helps me put extra effort into writing/analysis, attain extra individuals, and helps my crippling chocolate milk habit. Assist me democratize an important concepts in AI Analysis and Engineering to over 100K readers weekly.
PS- We comply with a “pay what you possibly can” mannequin, which lets you help inside your means. Check out this post for more details and to find a plan that works for you.
If you liked this article and wish to share it, please refer to the following guidelines.
That’s it for this piece. I admire your time. As all the time, for those who’re inquisitive about working with me or testing my different work, my hyperlinks can be on the finish of this electronic mail/publish. And for those who discovered worth on this write-up, I’d admire you sharing it with extra individuals. It’s word-of-mouth referrals like yours that assist me develop. One of the best ways to share testimonials is to share articles and tag me in your publish so I can see/share it.
I frequently share mini-updates on what I learn on the Microblogging websites X(https://twitter.com/Machine01776819), Threads(https://www.threads.net/@iseethings404), and TikTok(https://www.tiktok.com/@devansh_ai_made_simple)- so comply with me there for those who’re inquisitive about maintaining with my learnings.
Use the hyperlinks beneath to take a look at my different content material, be taught extra about tutoring, attain out to me about tasks, or simply to say hello.
Small Snippets about Tech, AI and Machine Learning over here
AI Newsletter- https://artificialintelligencemadesimple.substack.com/
My grandma’s favorite Tech Newsletter- https://codinginterviewsmadesimple.substack.com/
Take a look at my different articles on Medium. : https://rb.gy/zn1aiu
My YouTube: https://rb.gy/88iwdd
Attain out to me on LinkedIn. Let’s join: https://rb.gy/m5ok2y
My Instagram: https://rb.gy/gmvuy9
My Twitter: https://twitter.com/Machine01776819
