Loading...
RightNow Agent performance data below is historical evidence from DAILY-USE.md and the shipped profile comments. Every value is tied to its recorded date and condition. None is a general guarantee or a current full-suite result.
These measurements used qwen3-8-flash-next through RunInfra with a reused warm daemon.
| Measurement | Recorded value |
|---|---|
| Pre-turn session open and prompt assembly | 1 to 3 ms |
| Model call for a trivial turn | 1,054 to 1,591 ms |
| Client spawn, connect, and exit | about 400 to 600 ms |
| End-to-end trivial headless turn | 1,551 to 2,261 ms |
| One-line bug fix | 17.7 seconds |
| Model portion of that fix | 14.7 seconds across 9 calls |
The September 1 record attributes 83 percent of the one-line task to model time.
| Condition | Recorded observation |
|---|---|
| Cloud profile, headless create-file, daemon reused | Session open measured 105 to 115 ms instead of 631 to 728 ms. |
| Three headless create-file runs per dynamic-tools setting | 1,976 prompt tokens with dynamic tools instead of 9,676; model call 861 to 1,051 ms instead of 1,431 ms, about 35 percent faster; all runs produced correct files. |
| Simple task record later identified as Nemotron-specific | About 1.3 seconds end to end, about 320 ms harness time, about 7 ms agent work, and the remainder at the provider. The same record later documented file-corruption failures for that model. |
| Provider-managed prompt cache observation | About 38.5 percent hit at steady state and about 30 percent cost reduction, with no affinity or cache-key header. |
| Five-shape model battery | Deepseek-v4-flash recorded 5 of 5; the September 1 shipped-profile sweep separately recorded qwen3-8-flash-next at 5 of 5 and about 1.5 seconds. These are model-specific historical observations. |
A project-build turn recorded 120,019 ms wall time and 96,873 ms of model time across 12 calls. The then-current 120-second limit stopped the turn during a file write, and the run did not claim success.
The shipped profile now uses a 600-second time budget, 24 base model calls, and a 5,000,000-microcent spend guard. The same historical run used five orientation calls and about 10 seconds of stream before its first action. The shipped profile now enables a bounded turn-1 workspace map, but no post-change speedup is claimed from that run.
echo hi command measured 309 ms end to end, with about 250 ms attributed to the post-exit quiet period.remote_fetch = false, two small-project runs measured 6,964 ms cold and 5,860 ms warm, with about 4.7 seconds of model time.Evidence limits: the September 1 daily-use record also contains an unverified interactive TUI, an old count of 82 failing unit tests, and no current macOS or Linux product run. Do not present those entries as current test results. The old protoc installation note is superseded by the vendored protoc fallback.