An agent that fixes a bug and opens a pull request is a catalogue item by now. What is new at Cloudflare is where it sits: on your domain, behind your Access policies, with your AI Gateway in front. The house is theirs, the keys are yours.
Fact
The facts: on 1 October 2026 Cloudflare announced Cloudflare OS, an "agent workspace" that connects to a company's data and systems. You connect a GitHub repository and ask the agent to explore the codebase, fix a bug, add a feature or open a pull request; Google Workspace gives it reading and writing of email in Gmail and access to Drive, docs and sheets; export goes to Excel, CSV, PDF, Markdown and HTML, with Word and PowerPoint promised "soon". The open source version is available today to deploy into your own account; the fully managed version has a waitlist and will be launched from the dashboard by choosing a domain, Access policies and an AI Gateway. No price is given. Clef are "decision models": they classify an input and return a typed answer with probabilities, which your code uses to route, escalate or hand to a human. Clef is built on Qwen 3.8-27B, Clef-flash on Qwen 3.5-9B, both with rank-256 adapters, Apache 2.0 licence on Hugging Face and hosted on Workers AI. Per Cloudflare, across 43 benchmarks Clef-flash answers in 38.8 ms median and 122.4 ms at the 95th percentile, Clef in 209.3 and 238.6 ms; Clef scores 98.47% on BFCL, 94.20% on Banking77 and 97.43% on CLINC150 with OOS. Reinforcement fine-tuning starts as a service with Cloudflare engineers at the customer, self-serve is promised later. AI Search is generally available from the same day, and billing starts on 1 November 2026: $0.75 per million tokens ingested, a further $0.50 per million for images, $2.00 per GB-month, $0.75 per 1,000 semantic queries and $0.10 per 1,000 full-text queries. Free every month on all Workers plans: 5 million tokens, 10 GB and 1,000 queries of each kind. New: image embeddings with Qwen3-VL-Embedding, OCR for scanned PDFs and files up to 10 MiB instead of 4.
The three pieces are one thought. The agent lives in OS, decides with Clef, remembers with AI Search. Cloudflare sells the room the model works in and meters the power at the door.
Clef is the most interesting part for a builder. Calling a large model for every "should I escalate" is slow and expensive. A small model that returns "spam: 0.97" does it in under 40 milliseconds, and you keep the threshold in code. The numbers are Cloudflare's, on their benchmarks. The weights are open. You try them on your own machine before you believe the percentages.
The cheap decision is the one you do not call the big model for.
AI Search is the other lesson. Beta means free and temporary. From 1 November there is a tariff and a free band that is enough for a small site. For a bigger one: 5 million tokens go fast if you upload scanned PDFs with OCR. Check how much you ingest now, while there is no bill.
Cloudflare OS is open source today, and the managed version sits behind a waitlist. Run it in your own account and you hold the keys and you pay for the mistakes. That goes for any room you let an agent into next to your email. Read what it can send before you connect it.
Official primary sources
Cloudflare - Cloudflare OS, 01.10.2026Cloudflare - Clef decision models, 01.10.2026Cloudflare - AI Search is generally available, 01.10.2026The visual is generated code-art, no third-party images.