Proficient in Agents

I started my career at Microsoft as a PM on the Office team. Office, alongside Windows, was one of the first products to ever reach 100 million humans, and this was before smartphones. The team took enormous pride in what they'd built: a new way for people to be productive at work. As I grew up, the entire world of office work reorganized itself around the PC, around email, Word, Excel, and Powerpoint. By the early 2000s, “Proficient in Word & Excel” had become shorthand for “can do knowledge work.”

We're on the cusp of the next transition in what it means to be proficient at office work. The evidence is everywhere if you know where to look. You see it in the concept of being ‘cracked’. You see it in the joy people are having with Open Claw. You see it every week at Agent Builders Breakfast as we compare the tools we use. A new proficiency is forming around building agents to get real work done, and this will be the future of office work.

What makes it feel like the Lotus 1-2-3 moment is that the artifact you create to do the work is just a document. That's all you need today to generate complex production software that accomplishes real work. When the language you author software in is the language you already speak, the barrier to participation collapses. Everyone can create software that does real work for real people.

To help orient yourself inside it, here's a framework I've been testing (send me your feedback!). Three kinds of agents are emerging:

Expert agents. The agent is the expert, and it acts as a mirror reflecting your skills. Claude Code is the canonical example. The best engineers I know are a hundred times more productive, writing a hundred times more code and, quite frankly, burning out fast. The context is too much. They're having to redefine what their patterns and practices need to be.

Task agents. Here, you are the expert of the task. You're not buying software; you're saying, “this specific task needs to get done this specific way.” We see this across enterprises, people using workflow tools to automate research, customer service, and process work. The tools are general purpose because of the power of MCP and CLIs, it's unclear where the lines will be drawn.

Domain agents. The category I'm most excited about, and the one without a clean enterprise correlate yet. Most people use Open Claw as a ‘personal’ domain agent today. My bet is many AI SaaS companies evolve here. A domain agent is one you hire. It's connected to all your tools, it carries your taste, and it operates autonomously on your behalf: a GTM agent, an analyst agent, a finance agent. What makes it different is that every system you rely on is embedded inside it, and you give it wide latitude to achieve your outcomes.

We're on the cusp of a new generation of productivity software. It will have billions of users. It will be as ordinary as a Word doc or a spreadsheet. And contributing to your company's AI brain will require expert use of these tools, the same way “proficient in Excel” defined a generation before us.

The resume line ‘proficient in agents’ hasn't been written yet.

We're the ones writing it.