Phone speed, measured.
Allocation, screen reads and complete tasks are different measurements. Here is what each takes.
Phone lifecycle
| Operation | Median | p95 | n |
|---|---|---|---|
| Allocate from pool | 104 ms | 128 ms | 5 |
| First screenshot after allocation | 829 ms | 1.02 s | 5 |
| Wake | 28 ms | 37 ms | 10 |
| First screenshot after wake | 385 ms | 559 ms | 10 |
| Read a screenshot | 108 ms | 121 ms | 30 |
| Fork: original paused | 339 ms | 419 ms | 3 |
| Fork: copy ready | 6.96 s | 8.12 s | 3 |
When can I see the screen?
Without the pool, cold creation measured 7.7–8.1 s in two KVM test runs. That sample is too small to report percentiles.
- New pooled phone
- About 0.93 s estimated
- Sleeping phone
- About 0.41 s estimated
These estimates add the stage medians above. They are not measured end-to-end medians. Allocation and wake timings alone do not include the first screenshot.
Read text before taking a screenshot
Appmap UI-tree queries measured 1.0 ms median and 3.5 ms maximum across 10 local reads in one KVM run. This is a local query measurement, not the agent’s full request time.
The older text-read path measured 1.99 s median and 2.14 s p95 across 10 reads. The paths and measurement boundaries differ; this is not an end-to-end speedup ratio.
Complete tasks with an agent
In a small internal Claude Sonnet benchmark, 34 of 36 counted attempts succeeded across 14 tasks. Median task time was 45 s. Tasks included web work, changing settings, setting an alarm and adding a contact.
These results describe the tested tasks and installed apps. Infrastructure outages were excluded from the counted attempts; 94% is not an availability or general app-success guarantee.
Scripted appmap results
Setting an alarm succeeded in 5 of 5 scripted runs: 3 steps, 1.09 s median and 2.09 s p90. Dark theme succeeded in 5 of 5 scripted runs: 3 steps and 0.85 s median. These runs did not use an AI agent.
What is still being measured?
App installation time, active-phone CPU use, capacity under load and task-level tail latency. Faster text reads and first screenshots are in development. Those targets are not included as results above.