I’m checking my own stack, after reading an excellent summary written by Shelly Palmer: How Anthropic Thinks About Agents, Workflows and Tasks. Her article is based upon a talk given by Barry Zhang/Anthropic at an AI Engineer conference in 2025. I’ll embed the video if you want to see his entire talk.

Summarizing the summary:
Resist building agents for everything. Scrutinize whether a task or workflow might be more effective, both in time and cost.

Here’s where I might be guilty…and may stay that way.

Being an enthusiastic explorer like you, of course I want to build everything AI. It’s the way I learn best. However, at the recent Deep Learning AI Dev Conference, I was checked by a Docker rep. Interested in their cloud containers, I explained the budtender stack, also divulging that cost was a concern. After a few pulls on his beard he said, “you know, you could build the budtender with very little AI or none at all.” He went on, “the tolerance and preference chips could be built as keyword database calls, with only the chat input handled by a model. It would be much faster and cheaper.” *Gulp* Wow, he’s absolutely right, Occam’s Razor to the gut.

BUT, there are other important considerations: mainly my education. I’m not interested in creating a glorified excel spreadsheet, I’m here to learn AI and how to build a process that will operate in production. And here’s another important distinction: the agent is conversational. I believe it will make the customer experience richer.

The takeaway here is not my decision to keep my agent; it’s the Pause. I’m reasoning through and making deliberate choices.

Heeeere’s Barry!