Stop building one giant agent. Build a dozen small ones.

GitHub's COO hasn't seen AI wipe out entire job functions, and neither have we. What actually works is narrower: small agents that each do one thing, that you can actually maintain.

Everyone keeps talking about AI wiping out entire job functions. I still haven't seen that happen. Apparently, neither has GitHub's COO.

What does seem to work is much narrower. You take one repetitive task off your plate, train an agent to handle that one thing, and let it run. Then you move on to the next one. None of these are especially impressive by themselves. But they add up.

The era of the perfect mega-skill is ending

For a while, every other post was "go download this skill," the perfectly managed thing that runs your entire workflow end to end. One massive, beautiful, all-knowing system that stitches a dozen tools together and produces some mega output. Whoever built the best one internally became the resident AI influencer.

That version mostly doesn't exist for us anymore. What we keep reaching for instead are incredibly micro agents that do one thing for us very, very well. Not "produce the full report." More like "identify the single most important piece of marketing information in this pile." One job, done cleanly.

A dozen small agents quietly run your week

The magic isn't in any single agent. It's in the stack of them. One drafts the follow-up. One files the receipts. One watches a channel and surfaces only the thing you'd actually want to see. Each one is boring on its own.

A dozen small agents later, a meaningful chunk of your week is happening without you touching it. That's the part nobody screenshots, because it isn't a flashy demo. It's just work that stopped landing on your desk.

This is exactly the shape General Input is built for. Each of those agents is one workflow you describe to Geni in plain English, "every weekday at 7am, do this one thing." They sit side by side in a single workspace, each on its own schedule or trigger, and you add the next one whenever a new task starts eating your time. Nobody has to maintain a monolith, because there isn't one.

If you've never built one, start with a single agent, or grab one of the ready-made templates and point it at your tools. Resist the urge to make it do five things. The first running agent teaches you more than any course.

The real reason small wins: you can fix it

Here's the part that matters most. The trick is keeping each agent small and easy to fix.

When the work changes, and it always changes, you want to tweak one piece. With a mega-skill, weeks go by, months go by, things shift, and now you're trying to surgically edit a giant all-knowing system you stopped understanding months ago. You're screwed. With a small agent, you open the one thing, change the one line, and move on.

On General Input that change is usually a sentence. A small agent is a short plain-English instruction, not a tangle of code, so adjusting it means telling Geni what's different now, not refactoring a system. And because each one is cheap to build and cheaper to throw away, you're never precious about the agent that stopped being useful.

This is also why I'm not worried about AI erasing whole roles. The work that's actually getting automated is granular. Someone still has to decide which task to carve off next, what good output looks like, and where a human check belongs. That's closer to operating than to being replaced.

So don't go hunting for the one perfect skill. Build the small one. Ship it. Then build the next.

Frequently asked questions

Why not just build one agent that does everything?
Because the work changes. A few months in, a mega-agent becomes a giant system you no longer fully understand, and one small change means untangling all of it. Small agents stay easy to read, easy to fix, and easy to throw away.
How small should each agent be?
One job. It pulls the most important marketing signal, drafts one kind of reply, or files one kind of document. If you can describe what it does in a single sentence without an "and," it is the right size.
Do a bunch of small agents actually add up to real time savings?
Yes, because they compound. None of them is impressive alone. But a dozen of them later, a meaningful chunk of your week is happening without you touching it.

Start with one small agent.

Pick one repetitive task this week and ship an agent that does just that. Then build the next one.