Apologia Pro Small Agents

Alberto Sánchez

Oct 26, 2026

Share

Working with small agents (Luna/Haiku) is better than one would expect. Their limitations force developers to make smaller, more defined tasks. This allows for an easier flow state, more involvement in the process and less cognitive offload, while preserving the productivity of the agent era.

Cover for Curiosities of the Microscope. Anonymous, 1852

During these last years I’ve started to feel a disconnect. This era of cheap intelligence has made me realize two things:

In tech, there are two types of people: those that like coding and those that like the result.

I am, unluckily, part of the first group. It is indeed unlucky because cheap intelligence makes most coding a bit obsolete. We’ve all felt it: shorter deadlines, bigger expectations and smaller teams. In a way it’s good, things are moving faster and we are seeing a progress unlike anything before. However, there are some of us that miss the craft of coding.

A week or so ago I realized I was overspending on my job’s Claude Code subscription. Which means I had to survive a couple of weeks with just Haiku —what a pain—. At first it was a bit frustrating, I could no longer one-shot tasks and everything took more effort. But, after a while, I started to really enjoy it. For my personal stuff I use Codex, so I decided to give Luna a try. I spent a whole day just coding, without breaking my flow. It really bridged the gap: the concentration and control of the old times with the productivity of the agent era.

The first thing everyone realizes with small agents is that you need to narrow the scope. This might be tedious

It really forces to be more granular. You can no longer ask the agent to create a whole service, you have to define it step by step, with more back-and-forth.

But the great thing is that the back-and-forth is smoother. Small agents are faster and the tasks are more granular, so the answers are faster. With bigger models you spend some time crafting a big prompt (defining specs, etc.), then you send a huge task, wait a bunch of time, then review it. Maybe you try to do two tasks at the same time to avoid the downtime, but it’s easy to lose context and waste a lot of time and mental effort.

Because you are more involved, it’s easier to keep the flow state.

It also forces you to make more architectural decisions (the first lesson you learn is that small agents are pretty terrible at making these decisions): where does this config file go, what is the agreement on the API, what algorithm are we using here.

This forces you to be on the loop, you no longer feel like a passive driver.

It also forces you to keep in touch with coding. One of my big fears is cognitive offloading [1], losing the ability to code after years of delegating it to the agents. This way I’m forced to take most of the decisions, review all of the code and keep working the coding muscle.

Honestly, it is much more fun. I feel the efficiency bump of agents while keeping the fun I’ve always felt while programming.

Small agents force more granular tasks, with faster iterations.

Just Try It

I feel like coding is such a personal thing. Over these years I’ve met people that prefer to code sequentially, others go from general to specific, some need to draw on the whiteboard before any line of code. I feel like all these options are perfectly valid. In the same way, developers have started developing their own specific flows with agents: spec-driven development, custom skills, multi-agent orchestrators... and they are still all valid. But maybe this way of working with smaller agents resonates with you in the same way it did with me.

As Karpathy puts it, «You can outsource your thinking, but you can’t outsource your understanding». That is what I’ve felt working with smaller agents: I can develop faster without giving up ownership.