You should make an experiment; take someone who never used any LLMs or agents, and tell them to use it for the first time in front of you, and tell them to build something like a calculator program or whatnot. Bonus points if they're ICs or at least not-managers.
I think there is a lot us engineers take for granted, when it comes to communicating via text, how to state things clearly and what we think/reason when we read things. A lot of people don't have those "skills" innate, and the first time they use LLMs, they basically don't know how to interact with them, until they realize what they're able to do and not. Then they also learn what to say to steer the model into the right way, this is quite literally a "prompt engineering" skill they're now learning.
It's a big part of what, in my experience, is separating the very good engineer from the iffy one: Do you have a good mental model, and can you put yourself in the shoes of people sitting in a different mental model? It makes you a better dev, and even more so when it comes to AI tools, which have their own kind of alien brain.
I like to chat with Claude about how to approach a given problem, bring in extra context, etc, before even really drafting up a plan, while other people dive into implementation immediately and go on wild goose chases.
90% of the time we end up in the same place in roughly the same amount of time, and there are obviously tradeoffs to spending more time planning vs implementing. I'm oversimplifying as well.
Until now, the actual act of writing code: terminology, syntax, etc. was a significant hurdle, and that underlying mindset was a very useful, but missing in a surprisingly large number of developers, skill.
Now with LLMs doing the work of "translate this into code," increasingly the only thing that matters is that exact ability. And developers that don't have it or can't develop it won't be developers for long.
But - I don't know if it was April, or May - but very recently - the coding harnesses paired with decent SOTA models like Opus 4.8/GPT 5.5 - just started showing a lot more consistency, and completeness, and sometimes downright clever behavior - that they started to become way more useful.
Just one out of hundred+ examples - I gave Claude Code (Opus 4.8 High) a complex task that involved consul, vault - but I had neglected to give it sandbox permission to download from hashicorp.com. So - it created a entire test harness that simulated both the behavior of Vault and Consul - created all it's test cases, verified that they passed - and when I came back 40 minutes later said that it was all done.
It's test harnesses so accurately simulated the behavior of Vault/Consul - that on first try - no refactoring whatsoever - all of the protobuf/AESGCM/API behavior (that has varied significantly between versions) - worked.
This was something that would have taken me, someone super super familiar with the code and tools and APIs - a minimum of 3 solid days of work - and that would likely involve hundreds of attempts and refactors as I unwound all the weird encryption and packaging layers. It zero-shotted a full solution without having an API to test against
If these agents actually have an actual test-harness - It's honestly hard to imagine what they can't do - subject only to imagination and budget at this point.
Speaking personally - something changed Between January and, Let's say May - in which instead of seeing these things as mostly interesting technology demonstration, in which the flaws outweighed the benefits - I now genuinely think they are the future of programming. I'm dubious that I'll write much software manually in the future - beyond what I do for personal pleasure.
Some people "got" LLMs back in 2022, others needed it to evolve a bit.
It's not unlike computers. I started using them back in the 90s and absolutely nobody I knew was interested, while today everyone carries one in their pockets...
1) The data they have flowing through the system that enabled them to build composer (which is much better than stock kimi 2.5) and is presumably allowing the training of a new model on space Xs compute.
2) Cursors new 'github' replacement.
3) Enterprise sales/traction
If you look at all of these together, it's not implausible that they end up mostly 'owning' coding in 5 years time. If they replace GitHub with something more compatible with agentic coding and bring it into their whole ecosystem providing cloud and local agents, PR review and own frontier coding model.
It's specialised vs 'borg' isn't it. One way of thinking is that the world is owned by Anthropic/OpenAI and coding is just one of many things their model and software does. Another view is we have a 'coding with LLMs' company that specialises in this field of endeavour. Hard to say which wins, but I think they have a shot.
Personally my only objection to cursor is that it's more expensive. That's it, otherwise it is great to be able to choose say GPT-5.5 when I want to work on backend and Opus when I want to work on front end. Great to have PR review built in. If they were able to get composer 3 to as good as GPT5.5 / fable at the price of composer 2.5 they'd be winning on price again.
They really need to change their trajectory then?
And regardless being owned by xAI, a failed AI company which turned into a datacentre operator probably won't help them to achieve that.
> Hard to say which wins, but I think they have a shot.
The market for "coding harnesses" and "AI IDEs" is already oversaturated and they are effectively a commodity at this point, you can use any of them with any provider more or less interchangeably.
I think near unlimited access to compute is exactly what they need to train a frontier level coding model and serve it cheaply and profitably.
> The market for "coding harnesses" and "AI IDEs" is already oversaturated
I think my entire point was that it's not just a AI IDE. It's a coding focused model (currently Composer 2.5, soon hopefully something better), a Github Replacement, PR review/Bug Bot, Cloud Agents and so on and so forth. It's a ecosystem. An enterprise signs a MSA with you and gets everything they need all in one place.
Yes because Grok failed and they now have "unlimited" compute they can sell to other. I mean you are right that if they did X, Y and Z they could be very successful but their is no indication that might happen. In any meaningfully way seems like Cursor has peaked a while ago.
> An enterprise
Well either they are the type of companies which just buys whatever Microsoft is selling OR they let their developers to mostly pick what they feel is the best tool for the job on their won. I don't think there is that much in between (and its a cutthroat market e.g. GitLab)
> a Github Replacement, PR review/Bug Bot, Cloud Agents
Those things are a dime a dozen, you can vibe code them in weeks/months and there plenty of options on the market already. Well not Github of course, but there are various reason for that which have little to do with product quality and features (not that I think there are many companies which could build a meaningful GH replacement in a realistic time period despite its many flaws).
I just don't really see a huge income stream for dev tools companies (just like there never was) they can skim of something from the top by reselling AI models (generally at zero or negative margins..) but that's not the most lucrative business model when you have no real moot.
So what? How much money is that making?
the failure is in converting regular people into actual ai product consumers. Companies are realising that the money is not in regular consumers but in enterprise and they are not considering grok as a serious alternative.
if anything, the name, the branding and the x/twitter affiliation has hurt adoption from money makers rather than help it.
so yes, people know it, but no one is willing to pay for it
Does it mean they are out of the race? I have no idea, but things don't look great.
Yes and no. I've used a few different harnesses with closed and open models and there is definitely something going on that makes some harnesses work better than others. Many of the differences are hard to pin down and some are things people don't care about. But I wouldn't say they are commodified just yet.
1. Memory use. I have colleagues complaining that Clause Code uses several GB of memory. Meanwhile I haven't heard about that regarding codex or goose, or even opencode for that matter.
2. Suitability for local models. When you use Anthropic models, you use Anthropic as a provider. They can have software between the model and your harness that will fix issues with the model. One notable thing that even the best open weights models struggle with is broken tool calls. There is a lot that a harness can do to fix broken tool calls when working with a straight up ollama running a raw GGUF file.
3. Ease of use with non mainstream models. OpenCode has GREAT coverage of models/providers. Goose, less so as it relies on people to set up their own anthropic or openai compatability settings. e.g. Zed doesn't let you use Z.ai (which, if you speak British English, sounds ironic because "zed ai" isn't directly supported by Zed the editor).
4. Worktree support. Opencode and probably all the TUI harnesses works in a local directory - so you need the terminal to be in the worktree. Zed, however, works centrally on your git repo and tracks the worktrees so you can bounce around your work in a single window.
Of these, '2' is maybe the most important one but also the hardest to pin down as a feature. '3' is a one time cost. Of course '1' could be a blocker for someone using a macbook air or neo.
I'm not a gitlab user, just learning it, so I can't say how half baked they are or not.
At a high level though it seems like a huge step forward than GitHub
This is the same data they used to finetune Kimi K2.5 to make their newer Composer models, which benchmark substantially better than Kimi K2.5.
I've heard they also want to build their own base models, which will also benefit from their large amount of high-quality training data. Which will solve Grok's model quality problem.
This is all unsourced conjecture of course. But it's what I've heard.
So it's not really fair to call it "fine tuning", it's an important part of building a coding model in 2026, and cursor have done a pretty good job with Composer
Maybe because SpaceX paid with monopoly money (all stock deal)?
xAI overbuilt their data centers - they can't find paying customers for them, that's the reason they made deals with other companies like Google to use their own datacenters.
Cursor has the opposite problem of not having enough capacity. So this works well for them together.
Weather it's worth it - if you beleive that AI will solve every problem then having a piece of the pie early on might be worth it.
Remember how when google bought youtube for 1.65 billions people thought they are crazy? Or when facebook bought instagram.
60B is a crazy number but might be worth it for someone fighting for world dominance :)
xai is on the line to delivery capacity they already sold to Google and most analysts think they are 50/50 on actually meeting it.
the only proof they have capacity is that musk claims all the money they are burning is going to datacenters and gpu (mostly because if he put it on anything else the lie would be obvious)
Or are they paying for talent? It seems like xAI is sorely lacking in talent, most likely due to the CEO and folks' aversion to him. By throwing around some SpaceX monopoly money he can trap some talent with retention clauses and try to invigorate his failed AI business.
I think these are good examples: in both of those cases the buyer had a plan to monetize.
If you are a user of Cursor, expect to pay more for it or switch.
For anyone that doesn't get the reference, please start here [1].
1: https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
From a valuation model - $4B ARR with rapid growth, and the ability to shift traffic to internal models (honestly, massive amount of the time "composer" - their internal model is fine, and obviously going to get better). Say 17x Multiple which isn't unheard for a rapidly growing Startup with solid future structural profit elements (moving to internal model) - that gets you to $68B.
Being able to compare outcomes for workflows involving competitors will obviously be v v v v useful.
If you resell something worth $5 for $5 while having to pay for R&D and operating expenses that's not exactly comparable with a company that's selling actual products.
> Say 17x Multiple
On an extremely low margin business it is, yet again that wouldn't be the stupidest thing in today's market.
But there are many model agnostic harnesses out there: OpenCode, Roo, Cline, and many others. And even Claude Code can be setup to use non-Anthropic models.
Open routers prices are no different than cursors and you can use any harness you want.
Big brain, small brains? Hmmm
Too much friction though, with Cursor its out of the box.
We’re in the new era where startups boast about and bought based on revenue and not on just a number of users with unclear path to monetizing as it had been for the previous couple decades.
We can also note that we see Thrive Capital (Kushner) again in a win.
It can't as long as there is plenty of AI without it.
The real differentiatior is that if $60B today turns out to be all thrown away in a worst-case scenario, it would be easily more affordable and there would be less negative impact than $47B at the time if it was all thrown away on Twitter.
I don't see how these systems can have the ability to be effectively expressive about all of the minutia, and not have all of the various different possible expressions lead to vastly different outcomes.
For example - I might (real world example from this morning):
"Create a script that installs hashicorp vault and consul, store the data on consul. Then create ahelper script that will fill the vault server with sample data. Add HTTPS support. Now write a framework that reads and decrypts the encrypted data in consul. Support old (pre 1.3) and new (post 1.3 vault). "
That generates a 6 page plan using Opus 4.8 w/1mm context, including notes on what to prioritize, what format to create the scripts in, etc... (My cursor guidance already has a couple months of hints as to what I want in terms of scaffolding unit tests, canonical linux, performance, security, etc...)
That 6 page plan is the "Prompt" - but it's entirely generated by Cursor/Opus. It's there to tweak if you want to emphasize, or provide some taste - but, honestly - it probably does a better job than I would - so ~90% of the time I just accept the plan as is.
What actually matters is the ability to communicate well in general, not anything LLM-specific. Being able to state what you want clearly and unambiguously, and having a sense for what additional information you need to dump, even when the other side claims they already have everything they need.
Cursor does have it's own model - it's a heavily reworked version of KimiK2, called "composer" - that I use a lot of the time when I have fairly straightforward tasks that don't require a lot of exploration or independent thought. Lot cheaper - the Input/CacheWrite/CacheRead/Output costs of Opus 4.8 are $5/$6.25/$0.5/$25 per mm tokens, vs $0.5/-/$0.2/$2.5.
So your workflow now looks like mine except I prefer a different editor and only use the latest and greatest model so Cursor basically offers nothing over Codex.
I disagree about prompt engineering, but it's one of those things that probably varies because of what language you use, what problems you solve, and the degree to which you care about the output. Unless I'm writing tests, I keep AI on a very short leash because I'm writing critical code used by a very large number of users. I have noticed big differences in output quality depending on how I steer AI. Without steering, it will happily leave in dead code, change the use of variables so they need to be renamed, assume or fail to assume invariants, etc. As I said in another comment, I think we won't need to do that for very much longer, but right now it seems essential.
Over time - I expect Composer will be cheaper than Opus 4.8 - but the nice thing about Cursor - you can flick between models.
And (this is purely a personal thing) - I really like the extensive collection of "Plans" that cursor tracks - there isn't really a similar thing in Claude Code - but I really like the Claude.AI interface for everything else. It's also a much better general knowledge agent - the Cursor Chat interface isn't as nice.
Maybe you haven’t looked deeper into what modern Claude can do?
It's not perfect, btw - to some degree you are at the mercy of which models they support - currently only 27 from Gemini, OpenAI, Anthropic, Grok, and Kimi (Just K2.5) - presumably because they have commercial arrangements with them. The "Bring your own Model" model requires you plunk in your API key - which sucks. And only one at a time.
To the best of my Knowledge, Claude Code only supports one model at a time if it's not one from Anthropic (which will use the the entire suite of Anthropic Models depending on the task) - and you have to override it to a single model with an environment variable at startup - no ability to flick between models from task to task.
Depending on your workflow - you can save 70-90% on costs just by chosing a reasonable model for really extensive tasks that don't require thinking, max context, etc....
The end point may squeeze quality of operations at the subminute time span for ground control environment seriously launching Starship rockets one an hour, for example.
Some of the issues (off the top of my head):
- Note - that my "product" was about 3,000 lines of code - so tiny. But https://metr.org/ should give you some insight into the complexity the models are capable of.
- you have to be able to imagine the product. If I have the time, and energy, to imagine what I want - the model will build it. Here is an example of a much better programmer than I and something he wanted built - https://www.boatbomber.com/blog/claude-fable-5
- These are the first drafts. On average - any complex system needs about 10 years and at least 1000 active and enthusiastic about reporting users to really get robust code. Writing if via LLM doesn't (at least so far in my experience) help that much in reducing bugs if you were previously following any semblance of TDD. Lots of bugs in the code - the products you listed above have literally tens of millions of years of user experiences and bug reports that got them to where they are today. No silver bullet yet - just faster, less effort - and it enables non-technical people to create (still buggy) products.
What a world we live in - "dating oneself" is measured in weeks/months! :)
What you're describing seems like a workflow for building toys only. There's currently no reality in which someone would actually know what the y,z features are before making them. A plan generated in 5min would likely suggest a suboptimal solution compared to what a good solution would look like (which might take a year or two to figure out, for a human, so still a week or so for SOTA models if at all possible). Building something in golang is cute, but hard to be convinced until more novel applications are being generated from prompts.
The data submitted by Cursor's users tho, that seems to be very valuable.
Workflow in Cursor is actually awesome - I'm a little outdated in how I use it - I still establish goals/objectives, rather than managing the loop which does so - but if you can think broadly enough - I find it's pretty efficient.
Key things I like about Cursor (and I recognize I'm dating myself a bit here): - Plan Mode is really solid - I shift-tab, have it go create the plan using whatever insanely expensive SOTA model is available - I will usually spend 5-10 minutes on the Plan - review it, maybe even tweak it a little. (though 90% of the time it's fine out of the gate)
The nice thing about Cursor (and honestly Claude Code, Codex) - there isn't really any "prompt engineering" involved. You just say, "Go Build me x - it should have y,z features - and build it in golang for me" - and that's it - the 3-4 page Plan comes back - usually pretty credible - and then you click "build.".