Gus Spathis

The through-line

Gus Spathis argues that the economics and the shape of software are changing at the same time, and that entrepreneurs who only notice the first half will build the wrong thing cheaply. On the cost side, code creation has become so cheap that he treats it as disposable: a system that "a short five years ago would have taken a team of uh probably 50 developers well over a year to develop" he vibecoded over a weekend, and then vibecoded again "at least five different times", each with a different architectural approach, because he wanted to solve a different problem each time [2]. On the shape side, he insists the interaction model itself is being replaced: "We're moving into an era of agents. And you will see this uh I believe will take over the world" [1]. The two halves meet in his standing instruction to founders, which is not to build faster but to build differently. "I think entrepreneurs need to absorb what agents are and rethink their businesses and reimagine their businesses to leverage agents" [1].

Running underneath both is a hard-nosed distinction between something that demonstrates an idea and something that survives contact with users. He is enthusiastic about how far a non-technical founder can now get alone, and equally blunt that the finish line has only moved, not disappeared [2].

On what vibe coding is actually for

He starts from Karpathy's original framing, that it is a kind of coding where you "give into the vibes", "embrace exponentials and forget that code exists", asking for the dumbest things and going with whatever the AI does until the code grows beyond your comprehension [2]. Spathis reads that literally: the output is "essentially throwaway code. You're testing out different ideas, different approaches. Um, you want to make it so that it's really cheap and easy for you to throw it away and start a new" [2]. That is the value, and it is not confined to simple interfaces. He describes using it on a very complex project and rerunning the build repeatedly precisely because "the creation of that code itself was so cheap with today's tools that I'm able to rapidly iterate and run these experiments. And that's really what vibe coding is great for" [2].

He also places the current wave in a long lineage rather than treating it as unprecedented. Before AI, the same ambition went by other names: rapid prototyping, clickable prototypes, mockups, demos, wireframes, and it drove products from Apple HyperCard to Macromedia Flash to Visual Basic, which was originally conceived as something a non-computer-scientist could use to throw an idea together and test it. Endless money has gone into this problem, and in his reading these tools "always peak for a bit and then they they kind of die away" [2].

On why you cannot push a prototype straight to production

His answer to the obvious question is that "most vibe coding is a demo. It's not an actual product" [2]. What a prototype implements is what he calls the happy path, "the most typical way the entrepreneur thinks that um their product should be used by eager and um smart and folks who are willing to read documentation" [2]. Real users behave otherwise. They will not read help documentation, they "want to click buttons in random ways", they put more data into fields than the system supports, and the consequence is that "the happy path is rarely the normal path once users get it in their hands" [2]. The practical implication he draws is about timing rather than tooling: founders need to know where it still makes sense to keep going themselves and where it makes sense to call in an expert or hire internally [2].

On the tool landscape and how fast it turns over

He treats the market as a fast-moving horse race and tracks it accordingly. At the accessible end he groups Vzero, Bolt and Replet as the simplest tools, easy to use without being a computer scientist, and singles out lovable as "a brilliant piece of software" that "keeps getting better" and is starting to encroach on the mid-tier [2]. Six months earlier the leaders were Windsurf and Cursor, with augment code appearing on the radar; OpenAI's attempted acquisition of Windsurf became "a sort of a fiasco of epic proportions", and in his view Windsurf, a good product, has since lost many users, leaving Cursor in roughly the number one spot with its familiar editor, file list and chat window, model-agnostic across ChatGPT, the various flavours of Claude and others [2]. The emerging tier, for people comfortable with a command line, is Claude Code, OpenAI's Codeex and Google's anti-gravity, and he gives Claude Code the edge, particularly with Opus 4.5, while conceding the ranking "feels like it's changing every week" [2].

Knowing this list is not trivia to him, it is a procurement test. He urges every entrepreneur to get comfortable leveraging these tools, and tells founders hiring engineers or choosing partners to keep the names in mind, because "if you're working with someone and they're not leveraging these tools, um you're throwing dollars out the window that you don't need to and quite significant dollars" [2].

On one-shotting and using coding agents for non-coding work

The moment that shocked him most was a one-shot: a very long, detailed prompt into Claude Code with Opus 4.5, thirty minutes away for a coffee, and on return "it had literally coded tens of thousands of lines of code and I ran it and the first time I ran it, it worked perfectly. It was stunning" [2]. The input was two dense academic papers on architectures for very complex AI-based systems, and the system built from them worked, which he flags as something that essentially never happens with systems of that complexity [2].

He is equally interested in what these tools do outside software. If you have some command-line familiarity or will watch a couple of YouTube videos, he recommends downloading Claude Code, because "it's good at doing pretty much whatever you want" [2]. Ask it for a marketing plan and it will do "a ridiculously good job". His own firm uses it to analyse acquisitions: feed it the financial documents and a checklist, let it generate the questions, supply the answers, and it returns yes or no decisions on the various aspects of a deal [2].

On agents as the successor to buttons and forms

His definition of the era being left behind is concrete: interacting with technology "by clicking buttons, moving dials, um putting data into forms, and hitting submit buttons" [1]. Chat alone does not escape it, because you can ask questions and get answers but "you can't ask it to do very much" [1]. The step change is connection to external systems, so that a request like paying a bill, connecting to HubSpot to answer customer web chats, or connecting to Stripe to issue a refund actually completes. As he puts it, "the AI is the brains, but it can then do all of these things outside of its little world" [1]. His working definition of an agent is "an autonomous uh an autonomous software entity that has goals and gets stuff done for you. Uh it does it for you. It doesn't ask you to go do it", the software equivalent of hitting full self-driving in a Tesla and arriving without touching anything [1].

He ties this directly back to vibe coding: once your app can reach third-party gateways, "you're going to get another 10x value ad by the things that you want your app to do", and "you can build vastly more sophisticated things with very little extra effort" [1]. The near-term applications he expects are unglamorous, starting with customer support, and he contrasts the current experience of asking an airline for a refund and being told to call an operator with an agentic version that connects to the payment gateway and processes it [1]. The longer horizon, agents booking trips and handling email on his behalf, he labels the Jetsons view while insisting many of those things are not far away [1].

On reimagining the product rather than building the one the client asked for

His clearest illustration of what agent thinking replaces is a build for an entrepreneur who fields thousands of daily email inquiries about a constantly negotiated resource, where some senders are credible buyers and others are kicking the tires [1]. The founder arrived with a conventional specification: a website where people log in, create an account, and receive notifications. Spathis pushed back, "Hey, why don't we reimagine what what you want to build here?", and the founder resisted, sceptical it would work and wanting it his way [1]. Given three weeks, the team instead plugged into the inbox, scanned incoming mail, routed genuine inquiries to an agent that negotiated the terms back and forth with the sender until it reached agreement, then surfaced the deal for a simple yes or no, with a second agent writing up the contract [1]. The point of the story is the substitution, not the speed: "we went from creating a sa a very traditional SAS platform where you log in and you do everything in that old style to a new way of doing things using agents that's radically more powerful and far more efficient for both sides of that transaction" [1]. His closing advice follows from it, that founders who start thinking this way and build it into their businesses "are going to be the leaders of this next great wave" [1].

Takeaways

• Treat vibe-coded output as disposable. The value is that iteration is cheap enough to rebuild the same system five times with five different architectures to test different problems [2].
• A prototype implements the happy path, and "the happy path is rarely the normal path once users get it in their hands", which is why demos do not go straight to production [2].
• Judge engineering partners by whether they use current AI tooling; if they are not, "you're throwing dollars out the window that you don't need to and quite significant dollars" [2].
• Know the tiers: Vzero, Bolt, Replet and lovable for non-technical prototyping; Cursor now leading the mid-tier after the Windsurf fallout; Claude Code, Codeex and anti-gravity for command-line users, with Claude Code and Opus 4.5 currently ahead [2].
• Coding agents are general-purpose. Claude Code is used for marketing plans and for acquisition analysis, taking financial documents plus a checklist and returning yes or no decisions [2].
• An agent is "an autonomous software entity that has goals and gets stuff done for you", distinguished from chat by its ability to connect to Stripe, HubSpot, email and payment gateways and complete the task [1].
• Connecting an app to third-party gateways multiplies what vibe coding is worth, giving "another 10x value ad" for very little extra effort [1].
• Before specifying a login-and-forms SaaS product, ask whether agents could do the job directly; one such rethink replaced a planned web platform with an inbox-connected negotiating agent in three weeks [1].