kagentHow fast do your agents wake up, and what do they need?
Agents sleep when nobody uses them, and wake up on the next message. Describe your agents to see how long a wake-up takes, which part takes the time, and the CPU, memory and storage they need.
Every conversation is new.
- Download
- Wait for the disk
- Agent start
- Restore
Point at a bar to see each part.
To run 100 agents at once
The sum of the agents' limits, with the agents packed evenly and waking up at different times. Agents that sleep use no CPU or memory, only snapshot storage.
What a wake-up is made of
Download
A wake-up after the node dropped its copy fetches the snapshot from object storage and decompresses it. Snapshots of code compress to about a third.
Wait for the disk
Each wake-up writes and reads its snapshot on the node's disk. A slow disk adds time when its work takes longer than the download and the restore.
Agent start
With a Volumes snapshot the agent process starts again. With a Memory snapshot it continues where it stopped. More than 1 core does not make it faster.
Restore
Puts the workspace files, and the process memory if the snapshot keeps it, back in the sandbox. Below 1 core it runs within the agent's CPU limit.
Measured on Agent Substrate v1.0.0-rc2 with gVisor, on one node with 16 vCPUs, with a minimal Go agent that starts in almost no time, and object storage in the same region. The values are medians of the measured wake-ups. Your agent's own start time adds to a wake-up with a Volumes snapshot.