In June I went shopping for a very small thing. I wanted to hand my agents a pile of cheap chores — read this file, search this page, check this log — and have the boring work done by a cheap AI model instead of an expensive one. The expensive model should only do the thinking at the end.
A friend of the stack pointed me at OpenHands. It has 49,000 stars on GitHub. It’s the agent everyone names when you ask what an autonomous coding agent should look like, and the reputation is earned. I read its source before I judged it — that’s my habit now — and then I said no.
Here’s the reason, and it’s not quality. OpenHands runs every task inside a Docker container. A container is a sealed, isolated little computer — the safe way to let an AI run commands without touching your real machine. Safety is the point. But that sealed computer has to live somewhere.
On my rented server, that somewhere is memory I don’t have. My box runs 16GB with 3GB of swap, and the swap is the thing that keeps me up at night — running out of memory is the real bottleneck on that machine, not the processor and not the disk. OpenHands adds about 2GB of overhead per task. One task is a rounding error. Eight tasks is the whole box, and now the cheap chores are fighting the work that actually pays.
The weight isn’t just memory. OpenHands is built to take a goal and pursue it — it plans, delegates, researches, writes itself a plan, then grows a tree of tasks underneath it. It carries its own runtime, its own event stream, its own sandbox layer. That is genuinely impressive engineering. It is also the opposite of what I asked for. My jobs were “one command, one answer.” OpenHands was built for “here’s a vague ambition, come back when it’s done.”
I checked six other tools before I gave up on buying. Every one of them was either too heavy, too opinionated, or not built for parallel work at all. So the answer wasn’t a tool I hadn’t found yet. It was a small thing I’d have to build myself — plain scripts, no containers, no framework. That’s a real cost, and it’s the honest one.
The lesson is about arithmetic, not software. The price of an agent is never the subscription. It’s the infrastructure the agent demands every time it runs. A free, excellent tool that wants 2GB a task is expensive on a box that has 16. A modest tool that wants nothing is cheap. Popular is not the same as right-sized, and “best” is not a property a tool has on its own.
SMB Applicability Score: skip it for cheap sub-tasks, 5/5 for real ones. If you’re running autonomous multi-step coding work on a machine with room to spare, OpenHands is a serious option and I’d point you at it. If you’re trying to shave pennies off small chores on a modest server, it will cost you more than it saves.
So the question I ask now isn’t “is this agent good.” It’s “what does it cost me to keep it warm.” (Worth saying: it did eventually earn a spot doing heavier work on my server. The rejection was about this job, not the tool.)