On agents

When an agent is worth building

Agents are genuinely useful when they actually save you time or money. I absolutely do not mean save you effort. There are so many agents in the world that save people effort, which is useless, because it doesn't improve output. In fact, by removing effort and not saving time, all you're doing is making the person worse at that task.

Where most agents fail

Customer support chatbots are a clear example. Research assistant agents that just summarize also don't save time, because when it's time to actually cite or build off of another paper, you have to fully go through and understand it anyway, whether it's through AI-assisted guided learning or just reading the paper.

Why agentic coding, Zapier, and Clay work

Agentic coding tools are useful (duh), but Zapier also has an amazing agent. Clay has a great agentic system as well.

It takes so much time to set up an automation properly. Same thing with analyzing GTM leads in Clay. Having an agent that correctly converts vague user intent into structured insights is the most useful part of the agent.

The gate before building one

The thing to detect is whether the data is so unstructured or hard to look through for insights that an agent is necessary. For example, I built an agent on my personal website, and it was genuinely useless because it only answered a small subset of questions about my work and projects. Instead, I've grown it so that it is able to answer tons of questions about my opinions, experiences, and my life.

What's overhyped

It's when people make an agent that no one has tested or used. I fell victim to this many times, but I'm not someone who can be the arbiter of whether or not an agent is useful. But if you use the agent in practice and it produces useful results for people by saving them time and money, then it's good. Same as any product.