When Your First Direct Report Is an Agent

I'd been beating my head against the wall trying to build an impact dashboard for the executive director of a nonprofit I work with.

Airtable, hours of fiddling and filtering. At some point I just said "I'm done with this," opened Claude with the Airtable connector, ran a couple of read-only tests to make sure it understood the base, and gave it one prompt describing what I needed. Ninety seconds later it returned an interactive dashboard with the breakdowns the ED needed. Better than what I'd been trying to build. When I walked him through it later, the value was obvious to him too. You don't always know how a funder is going to want to slice your data, and being able to interact with it in natural language and get clean visuals back is a different relationship to information than most teams have ever had with their own database.

That was the warm-up. The thing that actually changed how I think about this work happened the next day.

The team has a recurring data problem. Several times a month, one of their employment partners sends updates about people in their program, and the updates arrive unstructured, in three different formats, and need to land in Airtable. Doing this by hand is the kind of work nobody wants to own. So I sat with one of the people on the team and we just talked through what would normally be involved if we were doing it manually. Open the message. Match the person. Find the record. Update fields A through F. Cross-reference a second sheet. Move to the next person. We counted the steps and got past twenty before we stopped counting. A few hours of focused work, every time the partner sent another batch. And it would compound over time.

I told her about that Bill Gates line, the one about how lazy employees are great because they always find an easier way to do things. We opened Claude, connected Airtable and started setting up the process together. We opened up Claude's chain of thought so we could read what the agent was actually doing and why. We talked through where human judgment had to stay in the loop and where we could let the agent rip. Where it had to pause for confirmation before pushing changes, and where we trusted it to just work.

Halfway through, something clicked.

No matter where you sit on the org chart, the moment you start working with an agent you're automatically leveled up. You have a direct report now. You're responsible for their work and their output. They're like a little Tamagotchi (where are my millennials at?). They need feeding, attention, instructions, or they don't function well.

I think this is such a cool idea, especially for nonprofits. Internal capacity goes up, of course. The data import that was going to be a few hours a month is now a five-minute job. But the thing that doesn't show up in the budget is that everyone on the team gets management reps. You build a leadership pipeline from the intern up to the executive director, accidentally, just by adopting agents in a coordinated way. That's leadership development that nobody had to schedule, budget, or send anyone to a conference for.

This is so obviously true once you're in the seat, and almost invisible if you're only using AI as a chatbot. The frontier labs aren't pouring money into making their chatbots smarter, because chatbots have kinda maxed out their utility. All the compute and training spend is going into how AI reshapes work, especially for enterprise. Across the frontier right now it's all about agents and how to get yours to actually do things for you. A chatbot is a peer you ask questions of. An agent is a worker you give instructions to. Different relationship, different responsibilities, different skill set.

At the end of the session we sat with it all for a minute. Sometimes part of what makes a session like this good is just the company. Sometimes it's enough to feel like you're not alone in the work.

I have a wide range of feelings about AI, from horror to delight, but one of my earliest and most durable ones is that I felt less alone with hard problems.

We also reflected on the nature of the work itself. It was distinctly repetitive. New data would arrive again, from different sources, in different shapes, and we'd need a way to handle it quickly each time. We reviewed the steps we'd taken with the agent. The role, the context, the steps, the judgment calls, who to escalate to. That's when those management instincts kicked in.

This is a job description. Agents are workers who can do almost anything, but they need to know specifically how to do it for you. In the agent ecosystem the artifact has a name: SKILL.md. The plain-language version is that it's an agent job description. Congratulations, you now have a direct report.

And we're not the only ones who'll need this skill. Anyone else on the team who touches this Airtable base can use it and ones like it, because once you create one skill you're start to create others. We could package them all up and pass it around and install our own plugin. Just like that, another management layer emerges. Someone's gonna have to make sure all these skills stay current and organized as the partners change formats, as the data model evolves, as new edge cases come in. Quiet new work that didn't exist last month.

If you run a nonprofit or a foundation and want to know where to start with agents, that's the answer. Find the soul-crushing twenty-step thing on someone's plate, sit with them, and write the job description together. You'll save the hours. You'll also walk away with a teammate who just learned what it feels like to manage someone. That's the part nobody is pricing in yet.

Wanna learn more about how we're building team, people and products with AI? Book time with me.