SaaS Midlife Crisis: In-House Agents Flop, and We're Too Proud to Use Someone Else's

Just got back from SaaStr. Biggest takeaway: everyone's racing to build agents, and they're all racing in the exact same direction — nobody wants users leaving their walled garden.

But it's not about churn. It's about context. Every time a user steps outside your app, your agent gets one less piece of information about them, and gets a little dumber. Retention isn't a renewal metric anymore. It's a feeding schedule.

Keep the user, feed the agent

So every app is bloating up like it took growth hormones. Marketing module, customer support module, analytics dashboard, tool integrations — they all want to swallow up your full eight hours. It looks like product expansion. It's actually the agent being hungry. It needs to watch you. The more it watches, the more coherent it sounds.

I feel this in my own work too. If something can live in one repo or one system, I refuse to split it. Splitting is "cleaner" — clear boundaries, single responsibility, every textbook tells you to do it. But agents don't care about textbooks. The agent doesn't know there's a helper sitting in the next project over. It doesn't know you defined that type last week. It'll just reinvent the wheel, every single time. Clean architecture is for humans. Agents want everything in one pot.

I wrote down a principle for this a while back: constraint engineering beats capability unlocking. An agent's intelligence ceiling isn't the model — it's how much the model can see. The same Claude, starting from scratch vs. starting with all your relevant memory, produces two completely different organisms. I went so far down this road that I now sync my own time-tracking data and financial transactions into a vault, and let the agent read up on how I've been doing before it opens its mouth. Mildly deranged. Works better than any "personalized" prompt.

The SaaS playbook for 2026 basically writes itself: keep the user → keep the data → feed the agent → keep the user. Beautiful loop. Closed system.

Most SaaS can't actually run that loop

The problem is, almost nobody can. Building a real agent requires serious volume, diverse use cases, and a permanent slot in the user's head. Most vertical SaaS doesn't have any of that — your users open your app for twenty minutes a day. The agent you bolt on top launches, runs for a month, and gets fewer DAUs than the "Contact Us" button on your homepage.

At which point the answer should be obvious: if you can't be the agent, be the outlet the agent plugs into. Become the tool. Let someone else's big agent call you. Their traffic, their data, your tool fee. Dignified. Profitable. Fine.

Being the tool feels like dying

Almost no one wants to. Because being the tool means cutting the cord with the user. From now on, the user talks to the agent, the agent orders your service on their behalf, and you never see the user again. For a CEO who's spent a decade staring at user growth curves, this feels worse than dying.

So they get stuck in the middle. Can't be the agent. Won't be the plug. The product starts looking like a middle-aged man: can't climb up to the position above, too proud to take the job below.

Twenty years later, back to square one

The funniest part: twenty years ago, SaaS's whole origin story was "do one thing well, let others call your API." Now that someone's actually trying to call them, suddenly it feels insulting.

History's a loop, apparently.