Every day, I open Twitter or the New York Times app and I see that a new software company has been wiped by the latest model release from our favorite frontier AI labs. The most recent example that comes to mind is ServiceNow, a company that builds enterprise workflows for IT, customer experience, and employee engagement. While recent earnings hit on conflict in the Middle East, the stock definitely took a tumble as well on the idea that the tool could be "vibe-coded." Before I continue, I'd also like to add that I am well aware that this discussion of a SaaS-pocalypse has happened many a time, but recently I've seen talk of it dim down a bit, which is why I wanted to share some ideas.

First of all, think about where coding agents succeed. Closed-ended tasks, such as SWE-bench and tau-Bench are the two that come to mind. Having worked with SWE-bench closely, I can attest that the benchmark tests coding ability via things like tool calls and the ability to write patches. However, it doesn't come close to the actual building of software. SWE-bench does nothing to the end of verifying that a database can handle production read-writes, scale up and down for capacity, schedule jobs and monitor processes, and understand customer needs.

I am not dumb enough to not know that agents can do all of these things. In fact, agents can do a lot more than just this. What is known, however, is that agents cannot do all of these things right now. If that was the case, every SaaS CEO would have replaced their whole employee set with agents; surely these CEOs aren't that unknowing about business and unit economics. That brings us to the question of why businesses aren't doing this. The answer, to me, is that these are just too intractable and difficult tasks. Agents can generalize and build new workflows for people on an individual basis, sure. Agents cannot build new frameworks and complex integrated systems for hundreds of thousands, or even millions, of users. The other thing to note is that while coding agents are good, they are evaluated on small-context tasks. SWE-bench stretches to a decent token size, but it's nowhere near the size of a real production database with 500,000 lines of disjoint, unported, un-linted code.

My argument is plainly that SaaS is good for now because it's old. Asking AI to build simple, horizontal platforms is easy, but vertical platforms are incredibly hard to replace with generic AI. Additionally, platforms that have accumulated a lot of data are impossible to replace. Using a model with memory and context is night-and-day as compared with models that are stateless and have no memory, even with the simple chat models that you and I use every day. Vibe-coded SaaS platforms are similar in this way. If you talk about a CRM with no intelligence and a single-digit amount of customers, an AI tool can replace that easily. If that's the case, your tool's market cap is also less than your net worth so there's nothing to worry about. However, if you have accumulated years and decades of customer data, selling patterns, behaviors, and other information on the people that use your platform, I refuse to believe that a vibe-coded tool can generalize and invent that intelligence. In fact, it's not a question of belief but rather of ability. Even with the most complicated agents, it would be impossible to generate that kind of authentic customer use data in line with spikes in traffic, geopolitics, and other events that drive and stall business processes.

I don't have any active investments in any of the companies I discussed above, but I am sure I may enter a position in a SaaS company that's been beaten down. In general, markets can be highly irrational, and this is most often underscored when I see 20-30% drops in a stock price overnight. Fundamentally, nothing in a business can have changed that much to justify a drop like that. Essentially, investors are saying that every 5 sales units becomes 4 with a 20% hit to a stock price, which makes absolutely no sense to me on a business intuition level.

If you think about anything, think about what tasks coding agents are really good at. Then, think about how well those tasks generalize to the real world. The Erdos problems, as a few folks (including Neel Somani - Go Bears!) have shown, are incredible cases where AI is able to solve math problems. However, AI has not solved the Navier-Stokes problem or come up with a faster implementation of Dijkstra's algorithm. Think about where we benchmark these models versus where they're actually used. That delta is how long you have to make money on SaaS companies.