OpenComputer previews serverless agents that resume with state intact
Digger's new platform lets developers test source-controlled agents locally, then deploy them on hibernating Linux VMs.
By Ryan Merket · Published
Why it matters
Long-running agents need durable files, state and execution environments. OpenComputer is betting that managed, hibernating VMs become the deployment layer beneath them.
Mohamed Habib (@motatoeshq), co-founder and CTO of Digger, opened an early preview of OpenComputer's serverless agents on August 4th, giving developers a managed path from local agent code to a persistent cloud deployment. Habib announced the preview in a post on X, showing the product's build, local test and deployment flow.
https://x.com/motatoeshq/status/2084502004618133724
Habib came to infrastructure through a less conventional route. He earned a PhD studying retinal image analysis before working on traffic routing, authentication, security and throttling for Amazon's Prime Video gateway, according to his professional profile. He later co-founded Digger with Igor Zalutski (@IgorZIJ) and Utpal Nadiger (@utpalnadiger), initially building open-source tooling for Terraform and other infrastructure-as-code workflows.
A full computer behind the serverless label
The "serverless" label describes how OpenComputer manages an agent's lifecycle. Each deployment runs on a full Linux virtual machine rather than a short-lived function. OpenComputer says the machine hibernates between invocations, preserving its memory, filesystem and working context, then resumes when another trigger arrives. Idle agents do not incur compute charges, according to the product page.
That architecture addresses a recurring deployment problem for agent developers. Agents may install packages, edit files, execute arbitrary code and continue a task across multiple interactions. Conventional stateless functions are designed around isolated requests with predictable runtimes. OpenComputer instead gives each agent a persistent workspace and a machine that can survive a process crash, disconnection or redeployment.
OpenComputer's sandbox documentation says each environment has its own kernel, filesystem, network and process space, with hardware-level isolation through KVM. The platform also supports checkpoints, VM forks and changes to CPU or memory while a machine is running.
Agents become source-controlled projects
OpenComputer packages an agent as a directory containing its identity, instructions, tools, service connections, runtime configuration, workspace and evaluation files. Developers can commit those components to a repository and review changes through the same process they use for application code.
The early-preview quickstart walks through a Gmail summarizer. A developer installs the OpenComputer CLI, initializes an agent from a template, defines its instructions and approval boundaries, connects a Google account, and tests the behavior locally. Running opencomputer deploy --alias production publishes an immutable deployment while retaining a stable agent identity. A later deployment moves the production alias to a new version.
That versioning model matters for agents connected to email, source code or business systems. Instructions and tool permissions can change behavior as materially as an application-code update. Keeping those files in source control gives engineering teams a record of what an agent was allowed to do at each deployment.
OpenComputer also assigns deployed agents permanent endpoints, allowing them to receive requests from webhooks, schedules and external services. The product is designed for jobs such as inbox triage, code review and background research that may need to pause, retain state and continue later.
Digger extends its infrastructure bet
OpenComputer is the latest step in Digger's move from managing human-written infrastructure changes to supplying infrastructure for autonomous software. In a March 15th essay, Zalutski argued that agent code combines the request-response behavior of an application with the isolation and arbitrary execution requirements of a continuous-integration job. His conclusion was that agents need a distinct deployment layer built around gateways, queues and dedicated execution environments.
The serverless-agents preview turns that argument into a managed product. Digger is betting that developers will want the convenience associated with serverless deployment while retaining a durable machine that an agent can treat as its own computer.
Digger announced a $3.6 million seed round led by Initialized Capital on June 16th, 2025. Zalutski said the round also included Olivier Pomel, David Cramer, Michael Grinich, Zeno Rocha, Ben Porterfield, Peter Zaitsev, Eran Sandler, Oana Olteanu and Palumni VC. Initialized said the founders drew on infrastructure work at Palantir, Amazon and Fitbit, while Nadiger had helped launch OpenTofu, the Linux Foundation-backed Terraform fork.
The preview enters its real test after deployment: whether persistent machines provide enough reliability and operational simplicity to justify another layer in the agent stack. OpenComputer's answer is to hide the machine's lifecycle without taking the machine away.