Oct 20, 2026 · local ai · tested at home
two coding agents on the same box, from 44 to 66 tok/s each in one evening
same machine, same model, one night of tuning: when two agents generate at the same time, each one goes from 43.8 to 66.4 tok/s, and four settings did most of the work
what i wanted to know
i run two coding agents in parallel on the same box, and at first it was honestly painful: as soon as one of them read a file, the other one froze at 2 or 3 tok/s. i wanted to see how far i could push the throughput of each one when they generate at the same time, without changing the machine or the model.
the test
always the same one, so i can compare one step to the next: two requests sent at the same time, 256 token answers, and the median of three rounds. on the side, i looked at what it gave in real life with two code reviews on an actual TypeScript repo, with contexts going up to ~180K tokens.
the progression
| step | per agent, two at once | solo |
|---|---|---|
| start: default batching, 4 slots, 220 W | 43.8 tok/s | 93.6 |
| 2 slots + PR #2985 (MTP drafts in batched slots) | 52.3 | 90.3 |
| + a friend's settings (prompt splitting, slot copies) | 54.7 | 91.7 |
| + Strata 0.2.0 beta (its tuning table for Ampere cards) | 61.1 | 107.6 |
| + both cards from 220 to 260 W and auto calibration | 66.4 | 109.6 |
in real use, each agent runs around 66 tok/s while the other one is generating too, and goes up to about 140 tok/s when it's alone, with 97 to 98% of the context reused from the cache.
what surprised me
the real gain isn't only in peak speed: before, an agent could drop to 2 tok/s while the other one was reading, and now the floor is around 15 to 20 tok/s, which completely changes how it feels to work with. i also found out that two fast slots beat four slow ones on this machine.
and most of the settings i tried did nothing: all the gain comes from the four things in the table, the rest of the night went into leads that didn't move the needle.
what i keep
two slots, PR #2985, Strata 0.2.0 and the cards at 260 W. since then i've undervolted them to get some silence back, at the cost of a few tok/s. to reproduce the setup, everything is in the strata guide.
what these numbers don't tell you
- each step was measured once (3 rounds), and the gap between steps 2 and 3 is within the noise
- the ~66 tok/s in real use is when both agents are generating; when the other one is reading a tool result, it drops (around 50 by feel)
- Strata PR #2985 isn't merged, it's a local build
are you an agent? this page in markdown · the measurements as json