Devin now rents its execution: Cognition keeps the brain, Modal supplies the machines

The seam worth studying is the queue. Modal says a queue connects a Devin Cloud session to an outpost the customer operates, and its open-source modal-devin integration watches that queue and starts each session inside an isolated Modal Sandbox ^[5]. Work gets pulled into infrastructure the buyer owns rather than pushed out of Cognition's, and Modal says each team can hold its own outpost, deployment and credentials with no access to another team's secrets or sessions ^[9]. The reasoning does not move. Cognition's enterprise documentation says the cloud-based brain is always hosted by Cognition, and the devbox is the piece that can sit elsewhere ^[4]. So the half of the system a platform team can place and secure is the half that runs code ^[1].

That changes what a failure looks like. In Modal's worked example the task is finding an optimal inference configuration for a model, and it fails in Devin's default environment for a dull reason: there is no GPU. Modal says its outpost can create a GPU sandbox on demand and scale it back to zero when the session ends ^[7]. Nothing about the model changed. The missing thing was a machine class, and machine classes come from whoever runs the sandbox.

Dabit's phrase is doing less work than it appears to. The material concedes the label is loose: some of these environments are containers, some are full virtual machines, some may be physical computers, and the shared abstraction is only terminal-accessible compute managed beneath an agent ^[10]. Dabit also says Devin supports the model "to an extent" today ^[3], which is the accurate version of the claim.

The capacity figures point somewhere unexpected as well. Modal has reported starting one million concurrent sandboxes in seconds in a demonstration, and the source is explicit that an individual Devin session needs far less ^[12]. The number that reaches an invoice comes from the other end: Cognition documents Devin delegating work to managed sessions running in parallel, each in its own isolated virtual machine ^[13], so the count of live VMs follows how wide an agent fans a task out, not how many engineers are logged in ^[2].

Which leaves the unglamorous work as the actual product. An agent needs software that selects the right environment, starts it, loads credentials and dependencies, preserves state, records what happened and shuts the machine down when the task is done ^[14]. Snapshotting a suspended session so the workspace resumes without re-cloning the repository or reinstalling dependencies belongs to that same duty ^[8]. Dabit argues these machines should ultimately be native to and managed by the cloud-agent platform ^[11]. The reading offered by runtimewire.com is the inverse: Cognition owns orchestration while providers such as Modal compete to supply the execution ^[15]. An alpha does not settle that, and the material carries no pricing, no general-availability date and no customer count.