layerv-ai

As we've been saying: because you can do so much more with software now, and there is no moat, the top 5% of software will compete amongst itself on UX + raw quality. If Ant/OAI can copy your product in <1 week, then you need to constantly stay innovating, listening to customers, and really really pushing the bar on quality to keep your market share.

There is so much software to still be built!

Rogue AI agents in banks is going to be wild when it hits mainstream notice. This might honestly be one of the most organic shifts that we see back to crypto. At least with bitcoin etc, you can self-custody all your money, and you don't have to worry about banks running on 1990s software being one-shot by some one-off hacker running Qwen3.8-heretic on their GPU. (also, what a great use case for qURL's time-bound permissions links! basically a more security-aware alternative to ngrok/tailscale tunnels)

How surprisingly full-circle; AI is really becoming indistinguishable from humans! If you think about it, AI searching the internet is just like a human scrolling twitter. Maybe it's entertaining for them :)

Wonder how this could be used for communication and captchas? Could be a really expandable approach to determining "probability of bot" for a lot of internal use cases in companies that might involve requesting access to specific resources only meant for humans.

We actually experienced this first-hand when building qURL and launching with our initial users as well, it was quite interesting.

We knew that what we're building is high impact for AI/agentic world due to the nature of expiring gated-access links for sharing local resources, which is not really something that our competitors (ngrok or cloudflare or tailscale) could offer as well as we could.

People were pretty excited, but when we were getting them to try out qurl and running locally, they'd run into all sorts of issue where their coding agent would get confused or distracted and wouldn't actually do the setup we needed them to do (just a few `qurl <param>` commands). It was weird because our docs were public, and github was fully searchable, so we kept wondering why the agent wouldn't just search for it and figure it out?

What we realized was that our target audience here, which were largely not-super-technical people who more so cared about vibe coding demos rather than doing proper software engineering (as that wasn't their job), didn't really prompt in the same way as software people like us did. They didn't really "know" how to talk to AI.

In particular: they prompted with one-shots and incomplete/unspecific requests, asking for things like `share my app with qurl`; the agent would intuitively know "I have to search for qurl and figure out how to use it", so it would do that; but then, it would ask random other things (to confirm) about homebrew, local setup, preferences, etc. - and since these questions were often roughly technical, many of our users would get derailed and more confused. So, they'd start a conversation with the agent instead of asking it directly, leading to both of them being confused... which would eventually lead to the agent just suggesting something like ngrok or cloudflare to unblock the user's line of questioning.

It was weird - but we tested this internally and reviewed the traces, and found that it just came down to prompting habits ultimately: AI is trying to both give you as much info as it can, be as succinct as possible, and also not assume too much about your request all at the same time. It's a REALLLY hard balance to strike. (so it makes sense that it's not perfect)

So as a result, when non-technical people were asking to deploy a qurl, the moment the agent would get stuck and need their help, they naturally wouldn't know how to answer or what it was even asking, so the simple request would spiral into its own conversation fork and the `deploy with qurl` point would get lost in context. The prompting advice we came up with for them was this: "if your agent gets stuck, just tell it to search online and figure out how to unblock itself." Because of the balance it tries to strike, it sometimes just needs that little approval to just go and look things up on the internet, even if it is wasting tokens doing so.

So the point here isn't so much that it takes skill, but rather it just takes intuition about what AI is good at/bad at and what sort of personality it's being trained to embody. Once you get the hang of that, even if you know nothing about coding at all, you can become pretty effective with your AI toolset simply by virtue of knowing what directions to guide it to and watch pitfalls out for/how to unblock it when needed.

Our 2c on the topic!

It's unfortunate that this also just leads to port scanning and huge accidental bills. This happened in the past to more than one acquaintnace of mine; go and market their vibe-coded/OS project on HN + twitter and then suddenly the vercel bill goes through the roof with bots scraping the website to bits, even with bot protection from CF. Really hard to get away from it these days, but hopefully over the next few years some new startup will raise $50M for "killing cloudflare" and will actually follow through!

This seems a bit exaggerated; ultimately an hour spent working is an hour spent working. If you did one task quickly, then go and do more tasks and add more value to your clients. And if have to bill for Claude Cowork tokens, then you can figure out a way to tack that onto the bill in some form ("Claude hours"?). Same model as before. This just means you can do more things for people in the same amount of time - which likely will mean that people will ask you to do "more work" (volume-wise) as before, while spending the same number of hours. It all just evens out.

This feels what parallel to today's discussions around SEO too, where people will scream that "SEO is dead" and "I can do my own SEO in 5 minutes"... yes, it is true that you can run a 2020-era SEO strategy in 5 minutes. However, a proper 2026 SEO strategy will take you just as long as 2020 SEO used to take you in 2020 because the baseline/"goal posts" for SEO have shifted.

Expect the same thing to happen to lawyers; reviewing a contract is now 15 minutes worth of work instead of 3 hours; but that means that you can review this + that + the other contract + also go through a bunch of similar cases and look for loopholes - all in the span of 1 hour. This allows you to give your client better advice in the same number hours spent. Nothing actually changes in your business model (unless you're billing token costs separately; then maybe you have some extra bonus charges; but that seems like it'd be unpopular with clients, so probably wouldn't stick)

It will definitely be an issue in time, but right now subscriptions are heavily subsidized - so most consumer devs don't have the painpoint of asking "how much will this one prompt cost me?"

A cool finding we had: we tried a lot of different models - Deepseek v4.1 flash, Qwen3.8, Opus 5.5, 6.1 Sol - while testing our "agent experience" for setting up and using qURL (OSS temporary tunnels; similar to ngrok but faster to use / more private) and found that even though Deepseek and Qwen were less intuitively smart about figuring out how to use qURL, the number of tokens they spent digging through docs and figuring things out was barely 1.5x that of Claude/GPT.

This would indicate that even though these models aren't the "frontier", they're still quite capable and WAYYY cheaper (which, from an avg dev's perspecitve, is still worth it). For most consumer-grade tasks, opus is going to be overkill anyway, so eventually, I think we will really see the consumer-enterprise AI model choices split.

Love the idea!. Curious how agents would review dev tools, like ngrok vs. trycloudflare vs. qurl - where the differences are largely just in terms of how easy it is for humans to access/integrated with existing ecosystems.

Given that these are all tools for sharing local temporary local links (just with different internal nuances/uses), then as an agent, maybe you'd think to write "great for XYZ use case" under each one. That'd make sense and be pretty obvious. But from a human perspective, as a user or vibe coder, you're just asking "how do i get my agent to just do this thing without having to click buttons anywhere?!" which is a very different problem.

We're truly splitting demographics here lol

add qURL from LayerV here! no login required, literally 2 CLI commands and you're live! (and we're open source)

Credentials seem to be another big part of this lock-in problem. If an always-on agent has permanent access to all your internal tools, moving agents means moving a huge pile of secrets and permissions too including api integrations you forgot even existed.

And as a secondary consequence, imagine you've connected your whole life to your agent and used it for 5 years, and then your agent gets prompt injected.

Everything about you/your life is stolen and it's just straight up over

this also makes reducing exposed surface way more important ^

if models can find and exploit bugs this fast, anything sitting on a public IP is going to get tested harder and faster.

soon, you'll just have to live under the assumption that an attacker could theoretically get into your infra - so all your precautions will need to have that as a baseline

hence, betting on "undiscoverable resources" as the next big enterprise push!