Back to the journal

· Brantley Brumley II

Build the Door: Why Software Companies Need to Let AI Agents In

Today my Vivint system told me the batteries were low on several sensors.

What I wanted to do was tell my AI agent, "Check which batteries those are and order any I don't have." That's the whole job.

What actually happened is I walked around my house taking sensors apart. One of them needed a flat coin cell battery, which, shockingly, I do not keep in a drawer.

My agent can log into the Vivint website, but the site is mostly billing. It doesn't show which devices I have, and there's no API or connector I can hand my agent. So the smartest tool in my house sat there while I played electrician.

I'm not picking on Vivint. They're a security company, and they should be careful. But I keep hitting this wall, and a lot of software companies are about to hit it from the other side.

I'm early. Everyone else is coming.

I use AI way more than the average person. I run a pile of agents for email, scheduling, research, and everyday errands. I'm also Chief AI Officer at GenAIPI (the General AI Proficiency Institute), so I spend my days helping businesses figure out where this is going.

The most frustrating part of using AI isn't the AI. It's working around the companies that won't let it in.

Most people aren't there yet. They will be. And when they are, they'll pick the services their agent can actually use.

Quick translation for normal people

An AI agent is an assistant that does things for you. It checks, books, pays, and orders. To do that with a company's software, it needs a way in:

  • API: a door built for software instead of people. The agent asks, it gets a straight answer, no clicking around.
  • MCP (Model Context Protocol): a standard plug that connects AI assistants to apps. Anthropic released it in late 2024, and it's now run under the Linux Foundation with OpenAI, Google, and Microsoft backing it.
  • CLI (command line tool): text commands an agent can type. Agents are really good at these.

When none of those exist, the agent opens the website and pretends to be a person, clicking buttons. It works, sort of. It's slow, and it breaks every time someone moves a button.

The numbers are already showing up

Companies are building instead of buying. In McKinsey's 2026 State of AI survey, 32% of respondents said their organization decided not to buy at least one software product or feature because they could build it with AI coding tools. That's not a prediction. That's a purchase that already didn't happen.

Investors already got scared once. In February 2026, software stocks lost about $1 trillion in market value in roughly a week. The first fear Forrester listed was that "SaaS companies won't be the provider of choice for AI agents." Forrester also called the "death of SaaS" story overstated, and I agree. Software isn't dying. Software agents can't use is the part in trouble.

Agents are coming in through the window anyway. Amazon sued Perplexity over its AI shopping agent. In August, the Ninth Circuit threw out the order blocking it, reasoning that the users, not Perplexity, were the ones accessing Amazon. The case isn't over, but customers clearly want to send their agents.

Everyone else is building doors. The official MCP registry went from about 1,200 servers at the end of 2025 to more than 53,000 by the end of October 2026.

Even Satya Nadella, who sells a lot of business apps, said it: "I think the notion that business applications exist, that's probably where they'll all collapse, right, in the agent era."

Why some companies are behind

From what I see working with businesses, it's usually one of these:

  • They just don't want to build it.
  • Their data is the moat, so they hold it a little hostage.
  • They don't think it would help them. (It would.)
  • Their engineers still aren't using AI to code, because someone is afraid it'll leave big holes.

Meanwhile companies a hundred times their size, xAI and the other big AI shops, use AI every day and ship this stuff constantly. Size isn't the excuse.

"But security"

Security matters, especially now. But look at what happens without a door.

My agent logs in with my full login and clicks around as me. It sees everything I see. There's no "read only," no "just the device list," and no off switch except changing my password.

An official door is safer. MCP's spec is built on OAuth, the same "sign in with" approval you already use. The agent gets a scoped key instead of your password, and you can revoke it anytime. Linear even offers a read-only connection.

Slack did it in the right order. In 2025 it tightened API limits on third-party apps to stop bulk data grabs. Then in February 2026 it shipped an official MCP server that respects user permissions. Protect the data, then build the door.

What good looks like

I'm honestly not all-in on MCP. It can eat a lot of your agent's memory. Cloudflare found that exposing its whole API as regular MCP tools would take about 1.17 million tokens, more than most models can hold, so it rebuilt the thing to use about 1,000.

A lot of the time a good CLI is just better. Simon Willison pointed out that giving a coding agent GitHub's gh tool gets you all that functionality "for a token cost close to zero."

And the customer shouldn't have to understand the setup. The agent can do it. The best newer companies already put a "copy this prompt" box right on the homepage:

Paste it into your agent. It walks you through login and access. You're using it. That's the bar.

It's not just startups. I use Gmail with my agents all day, and it's great. QuickBooks is getting there too. Intuit has an official QuickBooks connector for Claude, and I've had good results connecting it through Composio.

API, MCP, CLI. I don't care which door you build. I care that there is one.

The part where your customers decide

When their agent can't use your service, people won't send you an angry email. They'll quietly move to a provider their agent can use. Or they'll have an AI build a rough version of your product over a weekend and cancel.

The winners won't have the fanciest built-in chatbot. They'll be the ones that let my agent find the coin cell and order it while I sit on the couch.

So here's my question: what's a service you pay for right now that you wish your AI could just use for you?

If you want to talk through what this means for your business, connect with me here or find me at conductingthecode.com.