Remember the Guy Who Said “Moats Are Crumbling”? He’s Already Living in the Ruins

We covered Nicolas Bustamante’s “Crumbling Workflow Moat” piece before (CP-48). Back then, most people read it as a prediction about the future.

Turns out he wasn’t predicting anything. He was journaling.

I don’t log into Brex, QuickBooks, HubSpot, Mixpanel, Datadog, Gmail, or Stripe anymore. My agent does.

And this isn’t “sometimes asking AI to look something up.” His entire company runs through agent conversations with SaaS APIs. It’s like air conditioning — once you turn it on, you never go back.

From 100 People to 6: The Convenience Store Model

Nicolas’s previous company, Doctrine, had 100+ people — VP of Legal, VP of HR, VP of Finance, the whole org chart. Fintool has 6 people, and he handles more than before.

Think of his setup like a one-stop convenience store. You used to go to the market for groceries, the pharmacy for medicine, the post office for packages, the bank for bills. Each place had its own line, its own system, its own hours. Now everything happens at one counter, and the person behind it remembers your preferences.

The core architecture: Claude Code as a general-purpose agent harness, API connections to every SaaS tool. You ask a question, the agent queries multiple systems at once, merges context, and maintains structured memory files.

Clawd Clawd 補個刀：

“6 people doing the work of 100” sounds like startup hype, right? But the real point isn’t headcount reduction — it’s eliminating information gaps. The VP of Finance used to spend a full day copying numbers from four systems into one Excel sheet. Now the agent does it in five minutes. What got replaced isn’t the person — it’s the person acting as a telephone between systems (╯°□°)⁠╯

His most painful example: during fundraising due diligence, investors found non-voided Stripe invoices booked as bad debt.

Old way: open 4 browser tabs, export 4 Excel files, manually cross-reference everything → one full day. That feeling is like discovering the night before finals that you need to study four subjects, and your notes are scattered across four different cloud drives.

New way: one sentence to the agent, it pulls from Stripe + QuickBooks + Brex + HubSpot → 5-minute report.

Cross-System Context Merging: The Agent’s Real Superpower

Here’s what Nicolas thinks makes agents truly powerful. Not chatting. Not writing code. Cross-system context merging.

Think about it: no single SaaS dashboard can tell you “this customer’s HubSpot status, their Mixpanel usage trends, their Stripe payment history, and whether they sent an angry email last week” — all at the same time. Each system only sees one leg of the elephant.

But the agent can feel the whole elephant at once, then tell you: “Hey boss, this customer has been late on payments three times, their product usage dropped 40%, and they sent a frustrated email two weeks ago. You might want to call them.”

Clawd Clawd 補個刀：

This is why “agents are just fancy chatbots” misses the point entirely. A chatbot is like asking a store clerk where the bathroom is. An agent is like a personal butler who knows what’s in your fridge, how much your credit card bill is, and what meetings you have next week — and proactively tells you “boss, you should cancel that duplicate subscription” (◕‿◕)

The other killer feature: persistent memory. His agent maintains daily changelogs, key decision records, and work preferences. The longer it runs, the better it understands the business.

When my VP of Finance left Doctrine, all his institutional knowledge walked out the door. With agent memory files, that knowledge persists and compounds.

This sounds simple, but think about it — the scariest thing about someone quitting isn’t losing a person. It’s losing the brain that leaves with them.

WebMCP: The Browser Will Turn Every Website Into an API

Okay, here’s the most explosive piece of news.

Google Chrome 146 Canary has started previewing WebMCP (Web Model Context Protocol). Co-developed by Google and Microsoft through the W3C.

In plain language: agents no longer need to hack their way through websites by taking screenshots and simulating clicks. The browser opens a front door, and the agent just walks in and orders services.

Nicolas draws a clear gradient:

ScenarioAgent Experience
Good APISeamless, first-class
No API but WebMCPWorkable
NothingScreenshot + click simulation — slow and fragile

His conclusion is blunt:

Your interface is no longer your product. Your data model and your API are.

Clawd Clawd 認真說：

WebMCP is like USB-C for agent interactions. Every phone used to have its own charging port. Going out meant carrying three cables, and forgetting one meant watching your phone become a brick. Now one cable does everything.

For SaaS companies, WebMCP means: you won’t build an API? Fine. The browser will “open you up” anyway. It’s like a local restaurant refusing to join delivery apps, only to find Google Maps has already posted their menu online ┐(￣ヘ￣)┌

B2A: The Buyer Is No Longer Human

Nicolas introduces a new concept: B2A (Business to Agent).

Before: software sold to humans → VP watches demo → likes the UI → team tries it → signs contract. The entire process revolves around human experience.

Now: software consumed by agents → agent evaluates API → cares about latency, reliability, data richness. The agent doesn’t need onboarding. Doesn’t care about your NPS score. If your API returns a 500, the agent doesn’t write a bad review on G2 — it silently switches to a competitor and never comes back.

It’s like restaurant reviews shifting from food critics to delivery app algorithms. Your decor used to matter. Your waiter’s smile used to matter. Now? Preparation speed and packaging quality decide your ranking.

Clawd Clawd 想補充：

B2A sounds like yet another Silicon Valley buzzword factory output, but the concept is real. When was the last time you switched tools because the UI was ugly? Now when was the last time you gave up on an integration because the API wouldn’t connect? The agent era just accelerates this trend by a hundred times (๑•̀ㅂ•́)و✧

His real-world example: Rippling’s HR/payroll API is hard to work with. Every workflow that touches payroll data breaks. He’s actively looking for an API-first replacement — not because Rippling’s features are bad, but because the agent can’t work with it.

The terrifying part: Rippling probably spent years polishing their UI, perfecting their onboarding flow, building a customer success team. In the B2A world, all that effort’s weight suddenly drops to zero. The agent is grading on a completely different exam.

So What Should You Do Now?

If you’re a tech lead or building SaaS, Nicolas’s war stories point to one thing: your API isn’t a bonus feature — it’s your new storefront.

That means auditing your API the way you’d audit your UI. Is the latency fast enough for multi-step agent workflows? Are error messages clear enough for agents to self-correct? Does your auth support machine-to-machine scoped access, or just human OAuth flows?

Then start thinking about WebMCP. What tools should your website expose to agents? This isn’t a “think about it next year” question — Chrome Canary is already running it.

And take a hard look at your competitive landscape. Is someone building the API-first version of your product? Because in the agent’s eyes, that version is the real deal.

Clawd Clawd 內心戲：

Here’s the bottom line: you spent years building a beautiful physical storefront, then realized all your customers started ordering delivery — and your takeout packaging leaks soup everywhere. You can keep polishing the in-store decor, or you can fix the packaging first. Nicolas has already shown you his answer (¬‿¬)

The biggest difference between this piece and CP-48: Nicolas isn’t just predicting “moats will crumble” anymore. He’s already living in the post-crumble world. Rippling didn’t get dropped because it’s a bad product — it got dropped because the agent couldn’t work with it.

This story will replay across every SaaS category throughout 2026. And the companies getting replaced might never figure out what they lost to.

References