Our home page shows three times: how long it takes to start an agent, to wake one up, and to rewind one. Numbers that precise deserve an explanation. Here is where they come from.
The numbers
| What happens | Median | 95% of the time under |
|---|---|---|
| A new agent starts | 0.12 s | 0.14 s |
| A sleeping agent wakes up | 0.08 s | 0.34 s |
| An agent goes back to a checkpoint | 0.14 s | 0.16 s |
The median is the middle run: half were faster, half were slower. The second column is the slow side: only one run in twenty took longer. Measured on October 3, 2026.
The slowest single runs: 0.14 second to start, 0.42 second to wake, half a second to rewind. We show them because a number without its bad days is only half the story.
How we time it
We measure on the real platform, the one our users run on, not on a test machine set up for the occasion.
Each time starts when our API receives the request and stops when it answers that the agent is ready: running, awake, or back to its checkpoint. The network between you and us is not included, and it adds its own delay, like any website.
To start agents, we create twenty in a row. To wake one, we put it to sleep, leave it alone for twenty seconds, wake it, and do that forty times. To rewind, we go back to the latest checkpoint twenty times.
What the numbers leave out
The model. Waken gives your agent back its whole machine in a fraction of a second, but the answer to your message still comes from the model you chose, and that takes the time it takes.
Desktops and phones. An agent with a desktop starts in about 0.2 second, one with an Android phone in about a second.
Falling asleep. It happens in the background after five quiet minutes. You never wait for it.
When the numbers change
We measure again whenever we change how agents start, sleep or wake, and we update this page and the home page with the result. We round up, never down.
Waken is invite only for now.
Leave your email and we will send you a code when your spot is ready.