For four years my relationship with AI was that of a power user: models embedded in our content and community operations, always with a human editing the output. This year that changed. I stopped just using AI and started building AI-native systems.
The unlock: agents with real tools
A model with a clever prompt is a party trick. A model wired to real tools, with access to your systems, your data, and actions it can actually take, is infrastructure. That is the shift of 2024. Instead of asking a model to draft something and pasting the result around, I now build agents that read the CRM, create the task, pull the analytics, and report back. The model stops being an oracle and becomes a worker.
Across our portfolio I have been making every product AI-first: a shared AI provider layer so any product can call models the same way, and the first experiments with products calling each other's capabilities instead of living as islands.
MCP: a standard port for agents
Anthropic just released the Model Context Protocol, an open standard for connecting AI assistants to systems and data. This sounds unglamorous and it matters enormously. Until now, every agent-to-system connection was a custom integration. MCP is the USB port version: build the server once, and any capable model can use your tools. I started building MCP servers for our own systems the week it shipped, and I expect this protocol, or something shaped exactly like it, to be how all agent infrastructure gets wired.
What I now measure
The question that used to drive planning was who to hire. The question I ask now is different: how many hours does this task take end to end, and does a supervised agent bring that number down? Some weeks the honest answer is no. But the trend line since January is unmistakable, and it points at a company design where a small team plus a fleet of supervised agents does the work that used to need a department.
Using AI since 2020 was the apprenticeship. Building it is the actual job. This is what I am doing now.