The moat of SaaS seems to have changed, but it hasn’t gone away, for several reasons.

1. Switching cost: If you’re using something like Jira or Linear today, even if you think it’s worthwhile to build and maintain a replacement, getting the organization to switch over to the new thing doesn’t happen overnight.

2. Deployment: I hadn’t considered this one until recently, but builders using agents to code often haven’t worked with infrastructure before. The trope is “hey, check out my new site: localhost:3000,” but the reality is that infrastructure is a new domain, even if you’ve gotten good at writing code with agents. I’ve seen some fascinating outages, like one caused by someone who was surprised when the agent called a delete endpoint on a database and nothing asked for confirmation. Many systems were built for the old world and are still catching up.

3. Who owns availability: Agent-coded prototypes are great, but real users and infrastructure come with additional responsibility. Who picks up the phone if a team in another time zone is having an issue with the software? In the case of production software, building the tool is just the first step. Most of the time, effort, energy, and cost comes from maintenance. If you haven’t maintained a production system before, you might not have even considered this, but it becomes real quickly.

4. Does getting all of the above right for a particular piece of in-house software actually cost less than paying for the SaaS product? It’s possible to rapidly build the software you want and need in a short amount of time, but are the team or individuals doing the work prepared for the real cost that comes after deployment? For most of the math to work, you either need inexpensive labor or an extremely expensive SaaS bill. If you’re an engineer reading this, it’s possible your salary costs more than the SaaS tool you’re attempting to replace for the whole company.

5. Tokens aren’t free. Maybe you’re all in on the dark software factory. You have agents monitoring and patching the systems. Agents taking internal feedback requests. Agents running on-call. All this inference isn’t free, and the cost is variable, compared to the roughly fixed cost of the initial development effort. With time, some margins will compress, and just the threat of building something in-house will allow larger orgs to negotiate better deals. But it seems we’re in for more of a reconfiguration than a reckoning.

---

So silly that Xcode certificates break when you still need to accept a new agreement and there is no way to just do it in the Xcode app

---

Every voice to text app I’ve tried is (or ends up) bloated, broken, or uses a cloud model. GPT-5.6 Sol creates exactly what I need in 3 prompts.