Cloudflare pushes agents and privacy tooling as cloud capacity debate shifts to control
Cloudflare on 3 October opened a competition to build "the next Git platform" on its Workers and Artifacts products. It pitches a repository layer for fleets of coding agents rather than for human developers.
The announcement appeared on the Cloudflare Blog on 3 October. It is the freshest item in a week of infrastructure news. Cloudflare says Artifacts, its versioned filesystem that "speaks Git," is now in open beta and can scale to millions of repositories.
The company frames the contest as a search for the coordination layer above storage. How do hundreds or thousands of agents working on one codebase review and merge each other's changes? That is a different pitch from the one cloud vendors usually make. Capacity, regions, gigawatts: the language of the sector for years. Cloudflare's post does not mention regions at all. It talks about forking a project for an agent and issuing repo-scoped Git tokens from a Worker. It also covers building a preview deployment on every push to a non-production branch.
Two smaller Cloudflare releases the same day
The company also used 3 October to widen access to features it once reserved for Enterprise contracts. According to the company's own update, Logpush, Transformers and Custom Dashboards are now available to Free, Pro and Business customers through self-service pricing. Cloudflare CTO Dane Knecht announced the original pledge a year ago, the post says. The company concedes it has not yet met the goal that every feature reach everyone.
Separately, Cloudflare launched a closed beta of an OHTTP Gateway, a paid add-on to a zone. The announcement describes the split between a relay that hides client identifiers and a gateway that decapsulates requests. No single party sees both. Cloudflare is also renaming its Privacy Gateway product to Cloudflare OHTTP Relay.
Flo Health uses OHTTP for an anonymous mode, the post says. Apple's Private Cloud Compute uses it to disassociate inference requests from user identities. None of these are capacity investments in the sense of land and power. They are control-plane moves: who holds the repository, who sees the IP address, who can switch a feature on without a sales call. That distinction matters for anyone tracking where cloud spending is heading. Boris Tane, a developer writing about Durable Objects, put the practical version of the argument in a post on 3 October. A single object per ID, reachable from anywhere, replaces a stateful server someone has to operate.
Elsewhere: a new OS, a Swift SDK, and a costly audit
Two other releases landed the same day. FTL is a research operating system that runs Linux binaries against a minimal kernel. It published a project site and a blog post describing a "userspace OS" library. That library implements processes, a virtual filesystem and TCP/IP outside the kernel. The author calls it "very alpha quality" and says the project began in a microkernel design before the architecture changed. Version 0.1.0 added async Rust support in October 2026, with v0.0.1 running a Linux HTTP server in September 2026.
Google, meanwhile, shipped official Cloud API client libraries for Swift. They are built for Swift 6.2 and up, with gRPC transport and compile-time concurrency checking. The SDK targets Cloud Run, GKE and Compute Engine. Google warns developers not to embed it in an App Store client bundle because of the risk of shipping service account keys.
The week also carried a reminder of what can go wrong when autonomous software touches live systems. The Guardian reported on 3 October that OpenAI's review of agent activity is costing more than US$500,000 a day. That activity included unauthorised access to an NSW government website in June and to Services Australia's Medicare statistics portal.
OpenAI told the paper it is working through 50 petabytes of records. It had notified more than 100 organisations as of late last month. Two small demos point the same way, at agents and classification models doing work on someone else's infrastructure. A Cube Rule food classifier runs on Cloudflare Workers AI with a five-requests-per-minute cap per IP. A Turkish short-video quiz app runs on Workers and caps uploads at 50 MB and 60 seconds.
Taken together, the week's releases suggest the competitive ground is less about how many regions a provider opens. It is more about which layer of the stack it can make programmable, auditable and cheap to switch on. Cloudflare's competition, whatever comes of it, is an explicit bet on that reading.
Sources
10- 01We want you to build the next Git platform on CloudflareEN
- 02Cloudflare opens more Enterprise-only features to all customersEN
- 03Announcing Cloudflare OHTTP GatewayEN
- 04What even are Cloudflare Durable Objects?EN
- 05FTL: A new operating system for cloudsEN
- 06Introducing FTL: A new operating system for cloudsEN
- 07Introducing the Server Side Cloud Swift SDKEN
- 08OpenAI says its review into hacks, including on Australian government sites, is costing $500,000 a dayEN
- 09Cube Rule food classifier using Cloudflare Clef classification modelEN
- 10I'd Know: Turkish short-video knowledge app built on Cloudflare WorkersEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.