DigitalOcean's Managed Agents held a running process in memory through a four-minute pause

The obvious version of this test would have proved nothing. The author, who posts on dev.to as remdore, pointed out that writing a file, pausing, resuming and reading it back shows only that a disk survived ^[15]. "A stopped container loses everything that was not written to a volume, so if that sentence is literally true it is a different kind of thing," the author wrote of the pause claim ^[3].

So the evidence had to live only in memory. A bash loop held a counter in a shell variable, incremented it once a second and appended it with a timestamp to /tmp/tick.log, a file the process never reads ^[5]. Had the session been rebuilt from a restored filesystem, the count would have restarted at one. The log shows two adjacent lines instead ^[7]:

``` 47 08:05:42 48 08:10:10 ```

Both came from PID 590, holding the same variable ^[7]. The gap between them is 268 seconds, in a loop that sleeps for one ^[1]. This is good test design, and it is cheap to copy: the whole morning cost about seven cents ^[4].

Getting the loop to survive took a second attempt. The exec channel into the sandbox kills its children when it closes, and the invocation that lived was this one ^[6]:

``` setsid nohup /tmp/tick.sh </dev/null >/dev/null 2>&1 & disown ```

Any background helper an agent starts over exec dies the same way unless it is detached ^[6].

Pause returned in 0.86 seconds and resume in 1.16 ^[8]. Creating a session took 15.97 seconds from the API call to READY, which the author called "slower than a container and about what a Firecracker-class VM costs you" ^[11]. The author reported that the paused session consumed nothing during the gap ^[16]. Sizes run from mars-1vcpu-1gb to mars-16vcpu-32gb ^[12].

I think the KVM finding explains the result. When the guest has its own kernel and block device ^[10], the platform can hold the whole machine, and nothing running inside it has to cooperate.

Fork was where the author's assumption failed. The expectation was a disk snapshot and a second sandbox with the same files, which the author said is what checkpoint-and-restore products usually do ^[17]. The live counter in every copy showed otherwise ^[14]. The two forks averaged about 15.5 seconds each ^[2], close to the cost of starting a fresh session ^[11].

The agent ran on DigitalOcean's own inference endpoint, and one DeepSeek v4 Pro prompt took 7.3 seconds with 15,793 tokens in and 91 out ^[13]. Inference and sandbox charges land on one account ^[13].

The evidence is one pause of under five minutes on a service in public preview ^[1], from a tester who could find nobody else who had checked the claim ^[18]. The run did not exercise open network connections or pauses measured in hours. The ticker's own timestamps jumped across the pause with no signal to the process ^[7], so I would expect an agent holding a lease or a token expiry in memory to need a clock check after every resume.