Abstract
Coding-agent harnesses are usually compared by model quality or task speed. This study asks a more physical question: what does each installed harness cost the local system while it starts and while it works? OpenCode, Codex CLI, Cursor Agent CLI, Pi, Grok Build, and fx were measured on a Linux workstation using fresh process trees, fixed coding tasks, and explicit solo and delegated arms.
1.47 MB
fx cold-start RSS
Lowest median startup footprint
240.6 MB
Pi light-task RSS
Lowest matched Luna-max solo task tree
12.92 s
Codex heavy CPU
Lowest matched Luna-max solo CPU time
220.4 MB
OpenRouter Grok RSS
Solo heavy-task supplement median
Quick read
The result at a glance
Bars show medians. Lower is lighter or less CPU-intensive.
Cold-start peak RSS
MBLower is lighter before model inference begins.
Heavy task: Luna-max solo RSS
MBThe closest apples-to-apples active-task comparison.
Heavy task: Luna-max solo CPU
CPU sTotal local user plus system CPU seconds; lower is less CPU-intensive.
Short version
There is no single winner. fx is the smallest cold start, Pi is the lightest matched Luna-max task tree, Codex uses the least heavy-task CPU, Cursor finishes the matched heavy task fastest, and Grok Build with OpenRouter Grok 4.6 is the strongest complete deployment observed in the supplement.
The question
The study separates cold startup from active deployment. A fresh CLI process answers how much local work a harness performs before inference. A task process tree answers what the harness, its tools, subprocesses, and model-driven workflow cost together.
That distinction matters. A small native binary can still become expensive once context, tools, and delegated workers are active. Conversely, a heavier CLI can complete a task quickly if its workflow is efficient.
What the measurements show
The matched Luna-max subset is the closest comparison between harnesses. Pi had the lowest peak process-tree RSS on both task sizes. Codex used the least CPU on the heavy task. Cursor completed that heavy task fastest. OpenCode produced the largest repeated task tree in this setup.
A separate OpenRouter supplement kept the Grok Build harness but routed it to x-ai/grok-4.6. All twelve supplement attempts completed, separating the earlier native Grok quota failure from the harness itself.
How it was measured
Twenty fresh version/help trials per harness measured startup. Three trials per arm/task cell measured deployment. Every task started from a fresh fixture copy and ran sequentially on one Linux x86-64 machine with four logical CPUs and approximately 15.5 GiB of RAM.
A 20 ms /proc sampler followed the complete descendant process tree. It recorded summed RSS, CPU seconds, process and thread peaks, context switches, and I/O. GNU time supplied a fallback for native processes that exited before a sampler tick.
- Light task: an eight-check retry-with-backoff maintenance issue.
- Heavy task: three independent implementation, security, and comprehension packages with 24 hidden checks.
- Solo arms disabled delegation where possible; delegated arms enabled each harness-specific mechanism.
What it does not prove
This is a deployment-envelope comparison, not a universal causal ranking of harness implementations. Remote models can choose different turn counts, tool calls, and child processes, and those choices change local resource use even when the model names look similar.
The task sample is small and synthetic, provider-side compute is invisible, and Grok native task repetition was quota-censored. The most defensible reading is practical: these are the resource envelopes observed for these versions, policies, models, and tasks.
| Harness | Peak RSS | CPU |
|---|---|---|
| fx | 1.47 MB | <0.01 s |
| Grok Build | 27.0 MB | 0.02 s |
| Codex CLI | 50.5 MB | 0.09 s |
| Pi | 157.9 MB | 2.51 s |
| OpenCode | 187.5 MB | 1.85 s |
| Cursor Agent | 206.4 MB | 1.76 s |
| Harness | Light RSS | Heavy RSS | Heavy wall |
|---|---|---|---|
| Pi | 240.6 MB | 535.8 MB | 1047.5 s |
| Cursor | 363.4 MB | 565.0 MB | 458.3 s |
| Codex CLI | 441.1 MB | 608.9 MB | 851.7 s |
| OpenCode | 742.8 MB | 899.6 MB | 836.5 s |
| Arm | Task | Peak RSS | Score |
|---|---|---|---|
| Solo | Light | 213.9 MB | 8/8 |
| Delegated | Light | 163.6 MB | 8/8 |
| Solo | Heavy | 220.4 MB | 23/24 |
| Delegated | Heavy | 307.7 MB | 23/24 |
Research note
This is a Katteke working paper, not an official arXiv submission or peer-reviewed publication. The source report, data, and replication notes are linked below so the claims can be inspected in context.
