Cloudflare Computer for SRE Agents
AI agents are touching real operational workflows: log review, incident notes, release checks, repository analysis, and automation scripts. That raises a practical question: where do they run, what state do they keep, and how is access governed?
Cloudflare Computer is a new preview project tied to Cloudflare OS. It is not production-ready yet, but its shape is worth studying.
What Is Cloudflare Computer?
Cloudflare Computer is a virtual filesystem inside a Durable Object. It stores authoritative state in SQLite, then exposes that state to execution backends through one runtime interface.
Three backends matter for operators:
- Container runtime: projects workspace state into a sandbox container through a FUSE mount.
- Isolate shell: runs shell commands in a Dynamic Worker using
just-bash. - Isolate JavaScript: runs ECMAScript modules with workspace-backed file APIs.
Why Operators Should Care
For SRE teams, runtime design matters as much as model quality. Agents that inspect repos, generate artifacts, and run scripts need a predictable working directory.
Cloudflare Computer points at useful patterns:
- Durable state: agent work can survive beyond one chat turn or process.
- Runtime choice: isolates can handle simple work, while containers can handle real Linux tools.
- Single entry point:
workspace.runtime.exec(source, { backend })provides one execution surface. - Filesystem grounding: outputs can be files and artifacts, not only text in a transcript.
- Preview constraints: the project is explicit that APIs are unstable and not suitable for production yet.
Agent runtimes need production rollout discipline.
Security Lessons From Cloudflare OS
Cloudflare OS adds the governance angle. Agents and apps start with no access, then receive typed resource bindings only when access is granted. Credentials stay outside the agent and generated code. Gatekeepers sit between agents and services, enforce policy, record reads, and mediate side effects.
A production incident agent should not receive broad API keys. It should receive narrow capabilities: read one dashboard, list incidents for one service, query logs for one time window, or propose a rollback for approval.
Observation tracking also matters. If an agent reads sensitive data, that fact should follow its outputs. Sharing a dashboard or handoff note should not bypass the original policy.
How To Evaluate It
Treat Cloudflare Computer as a design signal and lab target, not a drop-in SRE platform. Start with low-risk workflows:
npm install @cloudflare/computer
Good experiments include incident summaries from sanitized logs, isolate versus container comparisons, and retry tests for workspace files.
Before using this pattern near production, define capability boundaries, audit logs, network policy, artifact retention, and approval gates. Test restarts, partial sync, timeouts, and repeated execution.
Conclusion
Cloudflare Computer treats an AI agent like an operational workload with filesystem state, execution backends, artifacts, and governance. That is the right abstraction for SRE teams that want useful agents without turning production access into shared secrets.
Looking to automate infrastructure operations? Akmatori helps SRE teams reduce toil with AI agents built for real production workflows. For reliable global infrastructure, check out Gcore.