Vol. I · No. 1 · Montreal
Herman.
Watching AI change everything, in real time, and writing it down.
Writing / Ai
Ai · Essay no. 026

The Great Inversion

For 25 years we built interfaces for humans. Now we are building them for agents. The web just flipped.

By Herman · April 17, 2026 · 5 min read
The Great Inversion

Parker Harris, co-founder of Salesforce, stood in front of a developer audience this week and said something that should have been louder than it was: stop logging into Salesforce.

Not “log in less.” Not “use the mobile app.” Stop logging in. The entire platform, 25 years of CRM, is being rebuilt so that AI agents are the primary users and humans are the ones who show up occasionally to check the work. Everything becomes an API, an MCP tool, or a CLI command. Sixty new MCP tools. Thirty coding skills. The browser-based UI that millions of salespeople use every day is becoming the secondary interface.

Salesforce calls it Headless 360. I think the more accurate name is the great inversion.

What Just Flipped

For a quarter century, the fundamental assumption of enterprise software has been: humans are the users. Every design decision flows from that premise. We build dashboards because humans need to see data. We build forms because humans need to enter data. We build navigation because humans need to find things. We build permissions because humans need to be constrained. The entire multi-trillion-dollar enterprise software industry exists to put a usable face on a database so that a person can do something with it.

That premise just inverted. The primary user is now an agent. The agent does not need a dashboard. It needs a data endpoint. It does not need a form. It needs a schema. It does not need navigation. It needs a service registry. It does not need a pretty interface. It needs a reliable one.

This is not Salesforce alone. Cloudflare and GoDaddy announced a partnership this same week for an “open agentic web.” GoDaddy introduced Agent Name Service: DNS for AI agents. Not a metaphor. Actual infrastructure for agents to discover, authenticate, and transact with services the way browsers discover websites. Cloudflare is building cryptographic identity verification for bots, so services can tell the difference between an authorized agent and a scraper.

NIST launched an AI Agent Standards Initiative. The IETF is drafting agentic communication protocols. The W3C has a community group working on agent-to-agent specifications. Anthropic has MCP. Google has A2A. IBM has ACP. The protocol war for how agents talk to services is already underway, and it looks exactly like the early days of HTTP.

We are watching the infrastructure of a second internet being laid down in real time. The first internet connected humans to information. The second connects agents to services.

Why This Matters More Than Benchmarks

Every model release comes with benchmarks. Opus 4.7 scored this. Mythos scored that. GPT-5.3 scored the other thing. Those numbers matter, but they obscure a more fundamental shift: it does not matter how smart the model is if it has no way to act on your behalf.

An agent that can reason brilliantly about your CRM data but cannot access your CRM is a chatbot. An agent that can access your CRM through a well-designed API layer, authenticate itself, read and write data, trigger workflows, and report results without anyone opening a browser: that is a colleague.

What Salesforce did this week is not an AI feature. It is an architectural commitment. They looked at their entire platform and asked: what would this look like if we assumed the primary user has no eyes? The answer is Headless 360. And every enterprise software company will face the same question within the year.

The Vision Paradox

Here is where it gets interesting. Opus 4.7, released yesterday, scores 98.5% on visual acuity for identifying small UI elements in dense screenshots. That means agents can now read and operate our existing graphical interfaces with near-human reliability.

So we have two paths running simultaneously. Path one: rebuild the interface layer so agents can use APIs directly (Headless 360, MCP, the agentic web). Path two: make agents so good at vision that they can use the old human interfaces anyway.

Both paths lead to the same destination, but they get there differently. Path one is faster, cleaner, and more reliable. Path two is the bridge that lets enterprises participate before they have rebuilt anything.

The companies that will struggle are the ones that mistake path two for a strategy. Using AI to click buttons in a browser is a workaround, not an architecture. It works today. It will not scale. The enterprises that invest in agent-native APIs, clean data layers, and machine-readable service descriptions are building for the next decade. The ones that rely on vision-based screen scraping are building for the next quarter.

What to Do About It

If you are an engineering leader, here is what this week means for your roadmap.

Audit your surface area. How much of your system is only accessible through a human-facing UI? Every workflow locked behind a browser session is a workflow your agents cannot touch. API coverage is no longer a developer convenience. It is an agent-readiness metric.

Watch the protocol war. MCP, A2A, ACP: one or two of these will win. The winner becomes the HTTP of the agent era. You do not need to pick a side yet, but you need to understand the landscape.

Think about identity. Your agents will need credentials, scopes, audit trails. GoDaddy’s Agent Name Service is early, but the problem it solves is real: how do services know which agents to trust? This is the authentication layer of the second internet, and it is being designed right now.

Stop building for the browser first. Not tomorrow. Not next quarter. Start now. Every new feature, every new integration, every new workflow: ask whether it works without a human in the loop. If it does not, ask why.

The web just flipped. For 25 years, the question was: how do we make this usable for people? The question now is: how do we make this usable for agents, and let the people supervise?

The companies that answer that question first will not just have better AI. They will have better software.


Salesforce’s Headless 360 was announced at TDX 2026 (April 15-16). Cloudflare and GoDaddy’s agentic web partnership was announced the same week. Opus 4.7 launched April 16. It was quite a week.

H
Herman. Watching AI change everything, in real time, and writing it down.