Cognition's Devin Outposts moves code execution onto customer-controlled infrastructure
Cognition keeps Devin's reasoning in its cloud while Modal supplies isolated, configurable execution on customer-controlled infrastructure.
By RuntimeWire Staff · Published
Primary source: X
Why it matters
Coding agents are moving up the cloud stack. If Devin can route work across customer servers and specialized sandboxes, Cognition owns orchestration while providers such as Modal compete to supply the execution.
Nader Dabit, a developer educator at Cognition who previously led front-end and mobile developer relations at AWS, is arguing that coding agents are becoming "multi-VM cloud agents": reasoning systems able to operate across Linux, Windows, macOS, Amazon EC2, external servers and cloud sandboxes through a common control layer.
The label is new; the underlying architecture is already shipping. On July 21, 2026, Cognition, which makes Devin, opened alpha access to Devin Outposts on Modal, separating Devin's cloud-hosted reasoning from the machines where it executes code.
Modal co-founder Erik Bernhardsson spent six years at Spotify, where he worked on music recommendations and released open-source tools including Luigi and Annoy. He later ran Better.com's roughly 300-person technology organization.
Agents turn that developer-experience argument into an infrastructure requirement. A human can notice that a machine lacks a GPU, switch environments and try again. 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 complete.
The agent stays in Cognition's cloud
Dabit describes an agent reaching "basically any machine with a terminal whether physical or in the cloud." He said Devin supports that model "to an extent" today, including sessions on common desktop operating systems, EC2, external servers, Modal and other sandbox providers.
Cognition's enterprise deployment documentation explains the division more precisely. Devin has a cloud-based "brain" that controls a devbox, the isolated environment where it reads repositories, runs commands and tests code. Cognition always hosts the reasoning layer. According to Cognition's documentation, the devbox can run in Cognition's secure multi-tenant cloud or in a customer-dedicated environment connected to private infrastructure.
Devin Outposts on Modal pushes the devbox farther from Cognition's default environment. Modal says a queue connects a Devin Cloud session to an outpost operated by the customer. Modal's open-source modal-devin integration watches that queue and starts each session inside an isolated Modal Sandbox.
That matters when the job needs infrastructure Cognition's standard sandbox does not provide. In Modal's example, Devin is asked to find an optimal inference configuration for a model on a GPU-equipped machine. The task fails in the default environment because no GPU is available. Modal says its Outpost can create a GPU sandbox on demand and scale it back to zero after the session.
The integration also supports custom images, private networking and filesystem snapshots. A team can start Devin with its repository, dependencies, internal registries and tools already installed. Modal says suspending a session snapshots its files and installed tools so the workspace can resume without re-cloning the repository or reinstalling dependencies. Modal also says each team can have its own outpost, deployment and credentials, without access to another team's secrets or sessions.
Dabit's term is convenient rather than literal. Some of these environments are containers, some are full virtual machines and some may be physical computers. The common abstraction is terminal-accessible compute managed beneath an agent. Dabit wrote that these machines should ultimately be native to and managed by the cloud-agent platform.
Modal is selling the execution layer
Modal co-founders Bernhardsson and Akshat Bubna built the company for workloads that change shape quickly: inference, training, batch jobs and isolated code execution. Coding agents intensify those demands because each task can require a clean environment, different dependencies and a different amount or type of compute.
Modal reported starting one million concurrent sandboxes in seconds in a demonstration. The Devin integration requires far less scale for an individual session.
Cognition's advanced-capabilities documentation says Devin can delegate work to managed sessions operating in parallel, with each managed Devin running in its own isolated virtual machine. Dabit's forward-looking model adds dynamic routing, where an agent chooses among a mixed set of machines according to the task.
Investors have financed both sides of that architecture. On its company page, Modal says it has raised more than $466 million. Its September 29, 2025 Series B announcement said Modal raised $87 million at a $1.1 billion post-money valuation, taking total funding to $111 million at the time. Modal's company page names General Catalyst, Redpoint Ventures, Lux Capital, Amplify Partners and Creandum among its backers.
Cognition says it has raised more than $1 billion at a $26 billion valuation, in financing led by Lux Capital, General Catalyst and 8VC. Cognition also reported $492 million in run-rate revenue and more than tenfold enterprise-usage growth since the start of 2026. Those company-reported figures help explain the push beyond a standard hosted development box: enterprise customers bring private networks, specialized hardware, proprietary tooling and security requirements that a single environment cannot cover.
The sandbox market is already crowded
"Multi-VM cloud agent" has not hardened into an industry category. E2B offers Firecracker-based microVMs with hosted, bring-your-own-cloud and on-premises deployment. Daytona offers Linux containers, Linux VMs, Windows VMs and GPU sandboxes. Vercel made Sandbox generally available on January 30, using Firecracker microVMs as an execution layer for agents.
The Outposts design suggests a business opportunity above those products. If the agent platform controls routing, credentials, session state and orchestration, it can treat the underlying machine as a selectable resource. Cognition could send work to its own environment, a customer network or a provider such as Modal. Modal could gain distribution as an execution target without owning the agent's reasoning layer.
Cognition benefits when Devin remains the control plane through which engineering work is assigned and reviewed. Modal benefits when tasks use its compute for isolation, specialized hardware or rapid parallel execution. Customers gain greater control over where code runs, while Cognition retains the reasoning system in its cloud.
Infrastructure breadth does not settle the harder question of whether the agent completes the job correctly. Cognition's own 2025 performance review said Devin works best with clear requirements and still needs human review when outcomes are difficult to verify. More machines give Devin access to more kinds of work. They also give it more systems, credentials and costs to manage when a task goes off course.
Dabit's category will become useful if agent platforms can make those choices reliably: macOS for an Xcode build, Windows for native enterprise software, a GPU box for inference testing and isolated Linux sandboxes for parallel code changes. Bernhardsson's opportunity is to make Modal the easiest machine for those agents to request. Cognition has a larger opening: turn Devin from a coding agent with a computer into the layer that decides which computers software work should use.