A Post SaaS world

A Post-SaaS Age of Coding Agents#

The original SaaS playbook was simple: build a decent product, put a subscription on it, and watch net retention do its quiet compound-interest thing. That era didn't end overnight, but it is ending. Code-writing AI agents are making sure of it.

We're focusing building tools for the post-SaaS world, where the question is no longer "Which app should we subscribe to?" but "Which outcome do I need given uncertain and shifting realities?"

This post lays out the three trends we're seeing talking to customers.

Trend 1: Build-vs-Buy Has Flipped#

It used to be obvious: people bought SaaS because building custom software was prohibitively expensive. Today, code-writing agents are shipping production-grade code with unit tests, integration scaffolding, and migrations while you still forget to unmute on Zoom while you are talking.

Consider a recent public write-up: a team of 16 agents, and roughly $20,000 in API costs, wrote a 100,000 Rust-based C compiler from scratch, capable of compiling the Linux kernel, amounting to about 100,000 lines of code [ref]. Besides, the hype, the probably spaghetti code outcome, and the obvious advertising value for Anthropic, this is an undeniable glimpse of what is coming.

Which is why CFOs should be asking: _"Why are we paying Salesforce $500,000 annually for a clunky product, when we can replicate 80% of the functionality with Claude Code in few days?". Two years ago, that sentence read like financial sci-fi. Now, it's close to real.

In fact, horizontal SaaS is being repriced by the market [ref].

A toy economic model for the flip#

Define Agent Hour: the cost of one productive hour of an AI coding agent (API + orchestration + review).
Compare to Seat License: annual per-seat price for a SaaS tool covering a use case.
One-time Build Cost: Agent Hours needed to achieve an 80% replica of your workflow.
Tail Cost: ongoing tweak hours per month as your workflows evolve.

When (One-time Build Cost + 12 × Tail Cost) < (Seat License × Number of Seats), "build wins". The Agent Hour cost keeps falling, while the value of owning your own source code (and data gravity) keeps rising. The breakeven crosses earlier than most orgs expect. And once you've built it, you can plug in your exact economics, not someone else's proxy metric.

Thought experiments#

Your RevOps team asks for "a custom attribution model with offline event stitching." In old SaaS, you file a ticket. In post-SaaS, you ship it in a sprint with an agent pair and a human reviewer.
Your finance team wants rolling cohort LTV with confidence intervals. The agent writes the pipeline, the forecast, and the unit tests. A data scientist tunes it. Your BI vendor quietly raises their price.
You need a workflow unique to your pricing rules and sales comp structure. SaaS says "That's an enterprise feature; let's talk in Q3." Agents say "We deployed it to staging."

Trend 2: Complete Customization, Outcomes First#

If build-vs-buy flipped the front door, complete customization kicks down the back. The new winners deliver outcomes so tailored, and so economically aligned, that retrofitted SaaS can't keep up. Fraud detection. Regulatory compliance. Risk management. These aren't "features." They're moving targets embedded in domain physics, policies, and incentives.

We've been writing about this trends for months now, but people are too busy thinking about data centers in space:

Cracking the long tail of data science problems [ref]
AI breaking Wright's Law on cost curves [ref]
"Your data isn't as ready as your slide deck says" [ref]

The conclusion was always the same: the long tail is where the value is, and it's finally addressable.

Outcome-driven economics, not feature subscriptions#

You don't pay for "number of users." You pay for "basis points of risk reduced," "MW-hours of curtailment avoided," "failure probability shaved," or "alpha net of slippage."
You don't accept "generic best practices." You demand "your practices, but instrumented, simulated, and optimized."

Trend 3: AI-Native Companies From Inception#

Retrofitting / bolting-on AI to companies is difficult, since people resist with the "we have always done things this way", "how do we know it will work?", "don't change what's not broken". In reality they are asking "what is going to happen about me? You can't teach old dog new tricks". This is fair and revelatory of creative destruction even when applied to human experience.

Unfortunately for who resists AI-native is compounding. It's like Roko's basilisk without the existential Kafkaesque dread.

We structured our company for the agent era: fully remote, everything documented, everything modular, everything unit tested. The good engineers practices make agents 10x efficient. We consistently see 10x throughput when strong engineering discipline meets agentic development. Meanwhile, the "vibe coding" approach (paste a prompt, push to prod) does produce excitement and a massive number of incident reports.

The market will reward capital efficiency and AI-first architectures. A few practical questions teams should be asking:

Can we use AI agents to scale customer support, sales, or onboarding?
Can we build features 10x faster with Claude Code or Cursor, without making code shitty?
Can we switch to outcome-driven KPIs and instrument our decisions end-to-end?

A brief note on "vibe coding"#

There's a temptation to treat AI agents like a magical intern who never sleeps and always comments their code. Tempting, yes. Also: the magical intern will happily string together three incompatible SDKs and cheerfully pass all unit tests you forgot to write.

If you have worked with interns (I have for 20 years), you know interns are as enthusiastic as dangerous.

You need engineering discipline: AI-native means agents plus process.

Summarizing#

SaaS isn't "dead," exactly; it's just been repriced by a new labor market: code-generating agents plus a smaller, sharper human loop.

In that world, the seat license looks strangely archaic, like paying rent on a calculator. The future is "pay for outcome" (we like to think "pay for decision").