Working with agents
Pragma and Hermes Agent
The same small local model, the same graphics card and the same four requests, given to Pragma and to Hermes Agent. What each one handed back.
People ask how Pragma compares with Hermes Agent, the open agent from Nous Research. So we ran both on the same model, on the same computer, with the same requests, on 26 September 2026, and kept everything they made.
How we ran it
- The model: Qwen3.5 9B (Q4_K_M), on one RTX 5070 with 12 GB. The same file for both.
- The model server: llama.cpp, the same CUDA build and settings for both. Hermes needs at least 64K tokens of context, so it got 128K, the most Pragma ever gives this model.
- Hermes Agent v0.21.5 (24 September 2026), from its own installer, run as
hermes chat --oneshot --yoloin an empty folder, so it approved its own actions. - Pragma 1.3.51, a new bot in auto mode.
- The requests: word for word the ones in the comparison on our home page, plus five questions about facts planted deep in War and Peace. Nobody touched either run once it started.
The results
| Request | Pragma | Hermes Agent |
|---|---|---|
| Instagram post | Right size (1080 x 1350) plus a sharper 2x copy, 3 min 40 s | Wrong size (1080 x 1080), plain, 54 s |
| One-page website | Complete, 2 min 19 s | Complete, 50 s |
| 6-second video | Counts to 1,000 and shows the line, 2 min 18 s | Counts to 1,000, the line never appears, 9 min 52 s |
| War and Peace, 5 questions | 5 of 5, 12 min 30 s | 0 of 5, 2 min 32 s |
Hermes was faster on the post and the website. Pragma's skills check the work before the bot says done (in this run it rendered the page, and looked at frames of its video), and that takes time.
Instagram post
Pragma made the portrait size the request asked for, with real fonts and a layout. It also added two things nobody gave it, a 7:00 AM opening time and an @sunrisebakery handle, which a person would have to catch. Hermes made a square with default type on a gradient, and saved it to the Downloads folder instead of the folder it was working in.
Website
Both pages open in a browser and have a hero, a menu, the hours and the address. Call this one even; Hermes did it in under a minute.
Video
Both videos count to 1,000. In Hermes's, the words "loaves baked this week" never appear, although its reply said they do. It also left its working files in the Windows temp folder.
War and Peace
The book is about 750,000 tokens, six times what the model can hold at once. Five made-up facts were planted in it, so the answers cannot come from memory, and the questions describe each fact in other words, so searching for a question's words mostly turns up the wrong pages.
Pragma read the whole book in parts, each part against each question, and answered all five. Hermes searched the file 62 times, read parts of it 14 times, and said it could not find the answers.
What this does not show
This is four requests on one small local model. Hermes does a great deal we did not test, such as messaging apps, scheduled jobs and its own skills, and on a large cloud model the gap could look very different. Timings come from one run each.
If you run Hermes, you can use it inside Pragma too: it is one of the engines a bot can run on.