Software as an Agent

I've been building startups for two decades. The pace at which you can build in 2026 isn't like anything I've seen before. Garry Tan says some YC founders are hitting $10M ARR with teams under 10, and that for a quarter of the current batch, 95% of the code was written by AI. Andrej Karpathy said on No Priors in March he hadn't typed a line of code since December.

Speed is the ultimate weapon, and it's going to kill a bunch of SaaS startups.

Satya Nadella said "SaaS will collapse" in December 2024 on the BG2 podcast. By February 2026 a Jefferies trader coined "SaaSpocalypse" to describe hundreds of billions in software market cap getting wiped on the same thesis. Marc Benioff just shipped Salesforce Headless 360, which is literally the CRM exposed as APIs and MCP endpoints with no browser interface. His line: "The API is the UI."

The mechanic is straightforward. SaaS apps were built around a human navigating a UI. When agents do the work, the UI is overhead. You don't need a dashboard if your agent reads the data and reports back. You don't need a settings page if your agent hits the database directly. The whole infrastructure of "things to click through" was solving a problem that's gone.

Software as an Agent

If SaaS is dead, long live the new king.

SaaS sold you software you log into. SaaA gives your agent software it operates on your behalf. There's no admin UI. No settings page. No dashboard. You're already in Claude/Cursor/Codex with access to your code and your data. That's the interface.

SaaA enables your business to run on agents from your console or editor.

You can see this everywhere

Karpathy isn't typing code. Tan built gstack: 23 opinionated tools that turn Claude Code into a virtual engineering team with a CEO, a designer, an eng manager, a release manager. He's running it across 40+ repos. Cloudflare just shipped Agent Lee, an in-dashboard AI assistant they're explicitly positioning as the new way to use Cloudflare. The dashboard still exists for them, but the trajectory is clear: the chat is the front door, the GUI is the museum.

Founders I'm talking to who are building AI-native businesses run their internal "ops" through Claude/Cursor/Codex with direct database access. No admin panel. No settings UI. The CEO asks the agent for the metric, gets a custom query, moves on with their day.

The companies slow to see this are still designing dashboards.

SaaS-for-builders is the most fucked category in software

SaaA isn't a startup category. SaaA is what founders build for themselves on the way to building startups. None of these tools are businesses. They're the new floor of personal infrastructure that used to require a SaaS subscription.

Which means the consequence of SaaA isn't "new SaaA companies replace SaaS companies." The consequence is sharper.

Anyone whose business model depends on selling tools to people who can build is selling to a customer who can now ship the same tool in an afternoon. That's the gun firing out both barrels. AI agents make founders 10x faster (barrel one) AND make the things founders used to buy obsolete (barrel two).

Document collaboration. Issue tracking. Sharing gated files. Scheduling meetings. Recording walkthroughs. Tracking events. Storing snippets. Anything where the buyer is technical enough to swap a SaaS subscription for a few SQL tables and an agent prompt is in the blast radius. The product can't differentiate. The only question left is whether you got to your audience before they realized they could build it themselves.

And then it goes wider

The current blast radius is technical founders. The next is everyone else.

Right now Claude and Cursor live in the editor. They're builder tools. But the trajectory is mainstream. When agents are good enough that a small business owner who's never written SQL can still point one at their data and ask questions, the same dynamic hits SMB SaaS. Not in 2026. But the trajectory is set. The "normies need dashboards" defense has a half-life.

The companies that survive this aren't the ones with the prettiest dashboards. They're the ones serving customers who can't or won't run agents themselves, in markets where the work isn't a few SQL tables. Regulated industries. Operationally heavy businesses. Anywhere the value is in physical execution, compliance, or human relationships, not in the software wrapper around them.

That's part of why I went back into healthcare. The vibe-coded slop crowds into easy markets first. SaaA doesn't change which sectors are valuable; it changes how few people you need to attack one. And it picks the survivors of the next decade based on whether your buyer was someone who could have built it themselves.

Build your own

If you've been a founder long enough that you have an editor open with access to your data, you don't need to wait for someone to ship the SaaA tool you want. Describe what you want to your agent. It builds the thing.

The future of software isn't the next dashboard. It's the absence of one.

I've been building SaaA tools for myself this year. Sharing a few of them publicly soon. More on that in the next post.