I’m working on an open-source coding agent called smelt. I want it to be fast and lean, with all the features of the other agents. I’m biased, so I needed an honest way to check my progress.
I wrote a [benchmark](https://leonardcser.github.io/harness-resource-bench) that captures the cost of running an agent. I ran it on
[,](https://claude.com/product/claude-code)
[codex](https://developers.openai.com/codex),
[,](https://fx.sh)
[omp](https://omp.sh),
[,](https://opencode.ai)
[pi](https://pi.dev)and
. Here are the best results, including where smelt falls short.
Takeaways #
- Memory use differs by up to 74x between agents.
- The TUI adds as much CPU as the work itself.
- smelt has the cheapest engine, but the TUI adds 13 times more CPU.
How I measured #
- All agents get the same scripted model, so their replies and tool calls are identical.
- Each run uses a new container with 2 CPUs.
- CPU and memory are read every 100 ms, from outside the container.
- Each case runs 3 times. The charts show the median, and the whiskers are the lowest and highest run.
The benchmark is open source and I ran the following versions: claude code 2.1.319, codex 0.167.0, fx 0.0.16, omp 18.18.0, opencode 2.2.6, pi 1.5.0 and smelt 0.8.0. Newer versions may differ.
Memory varies 74x #
opencode takes 430 MiB directly after starting. fx takes 5.8 MiB. smelt takes 19 MiB.
The TUI costs as much as the task #
Drawing onto the terminal costs as much CPU as the task itself or more. codex, omp and opencode are exceptions.
A faster model makes it worse. From 30 to 1,000 tokens per second, claude code goes from 3% to 43% of one core, opencode from 4% to 48%.
Where smelt falls short #
The smelt engine uses the least CPU. 0.12 s is spent on the task. The TUI adds 1.5 s to it. That’s 13 times more. With the TUI, smelt uses more CPU than pi (1.4 s). The UI is where smelt can improve most.
Limits #
I wrote the benchmark as well as smelt, and I built smelt with these numbers in front of me. It’s a scripted model, so none of this gives us any idea of the quality of answers. It all ran on one Ryzen 9 7900X machine. The specifics of each agent’s setup are in the repo.
The full results have every chart, including idle CPU, startup and long sessions.
A note: I put real thought into these posts. Every word here comes from the human me. Thanks for reading.