Skip to content

About

Does a paused agent session really keep its memory? Measured on DigitalOcean Managed Agents.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

2 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

managed-agents-pause-test

What I actually measured when running a coding agent on DigitalOcean's Managed Agents (Harness Runtime), in public preview, on 29 September 2026.

The question was narrow: when the docs say a paused session preserves "processes, memory, and the workspace filesystem", is that literally true?

So I started a process whose counter exists only in memory, never read back from disk, paused the session, waited four and a half minutes, and resumed.

the counter across the pause

47 08:05:42     <- paused here
48 08:10:10     <- resumed here, same PID

The counter continued from 47 to 48 with the same PID. The only trace of the pause is a 4m28s hole in the timestamps.

measured timings

Forking was the bigger surprise. Two children created from the running parent both came up with the same process at the same PID, then diverged:

ticker PID counter shortly after
parent 590 244
child 1 590 231
child 2 590 228

Files

  • agent.yaml — the session manifest used for the pause test
  • locked.yaml — the same thing with an egress allowlist, which flips the sandbox to deny-by-default
  • tick.log — the raw counter log, including the gap

Reproducing

doctl harness-runtime create -f agent.yaml
doctl harness-runtime pause <session-id>
doctl harness-runtime resume <session-id>

Total cost of everything here, four sessions including two forks: about 7 cents.

MIT licensed.

Follow-up: checkpoint, fork three ways, roll back

An agent built a small web app inside a session. I checkpointed it, forked it three ways from that checkpoint, gave each fork a different brief, and screenshotted all three through doctl harness-runtime port-forward. Then I deleted the parent's app, killed its web server, and rolled it back.

one checkpoint, three futures

agent builds the app 91.6 s, 29,084 tokens
checkpoint the running session 25.2 s
fork x3 from that checkpoint 15.3 s
roll back after deleting the app 5.6 s

Every fork came up with the web server already running, same PID 593, and after the rollback the killed server was back at PID 593 too. The restored page was pixel-identical to the original except for 88 pixels in a CSS-animated status dot.

Checkpoints belong to one session. Rolling a fork back to its parent's checkpoint fails with checkpoint not found, so checkpoint the fork itself if you want to undo inside it.

same prompt, different app timings

About

Does a paused agent session really keep its memory? Measured on DigitalOcean Managed Agents.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors