Cloudflare (Workers AI)
Top signals
- #1WritingAgents can now create Cloudflare accounts, buy domains, and deploy8.0
- #2WritingAnnouncing Cloudflare K2: serverless event streams8.0
- #3WritingCloudflare OS: an open platform for agents, apps, and work8.0
- #4WritingCloudflare OS: your company’s agent workspace, managed for you8.0
- #5WritingIntroducing Clef: our open-source decision models, and new RL fine-tuning platform8.0
Agent answer
Cloudflare (Workers AI) has 3,801 loaded public signals: 850 hiring, 4 forks, 2,562 releases or model cards, 191 talking, and 194 repos. Latest signal: AI Search is now generally available. Data-business radar is currently scoped to frontier labs, so this category does not expose radar lanes. The standing analysis was generated with deepseek-v4-pro and 94 evidence refs.
has loaded 3,801 public signals
has hiring signal count 850
has fork signal count 4
has release signal count 2,562
Thesis
Cloudflare is converging its edge network, security franchise, and a new first-party AI layer into an “agentic Internet” platform play — with Workers AI moving from inference resale toward differentiated, self-trained decision tooling. The decisive signal is the October 1, 2026 launch of Clef and Clef-flash, “the first models trained by the Cloudflare Workers AI team,” open-sourced on Hugging Face under Apache 2.0 and paired with a new reinforcement-learning fine-tuning service W1. Rather than generate text, Clef emits per-option probabilities over a typed question schema W2, and is fine-tuned on Qwen backbones (Qwen3.8-27B / Qwen3.5-9B) W4. Around that core, the pack shows parallel investment in an agent runtime (interrupt/replay primitives in @cloudflare/computer P17, agent harness recovery P20), a near-daily runtime cadence (workerd P27), and hiring that pairs infrastructure scale-out with agentic-commerce and AI roles E9E42E43E10W6. Cloudflare's public framing extends the same bet to enterprise trust: an “agentic Internet” business-model thesis P7, SASE positioning aimed at “unmanaged AI agents” and MCP sprawl P5, and post-quantum, geo-fenced key management for sovereign accounts P28.
Signal desks
Hiring
- AI/ML buildout: Workers AI Engineer (Austin, $145k–$195k, hybrid) W6; Senior Machine Learning Engineer E10; team growth via Ensemble AI talent focused on “machine learning infrastructure and efficiency” W5.
- Agentic commerce: Software Engineer and Senior Software Engineer, Agentic Commerce (both In-Office) E42E43.
- GTM automation/commercialization: Principal, GTM Automation, RevOps to “build the AI agents and automation” for Sales/Marketing/Customer Success/Partner/RevOps (Austin/SF) P14E18; Senior Manager, Demand Marketing E59; Senior Events Manager (Austin) P11.
- Core platform/infrastructure: Senior Infrastructure Engineer, Storage Platform E9; Senior Software Engineers for Cloudflare Queues E5, Hyperdrive E7, Load Balancing E57, and Deploy at Scale E55E56; Senior Systems Engineers for Cloudflare Tunnel E34 and Basin Pipelines E12; Principal Software Engineer – Fintech E58; Principal Engineer E30.
- Network/datacenter ops: Data Center Network Operations Engineer across Atlanta, Austin, Denver, New York, San Francisco, Seattle, and Washington DC P10, plus Lisbon/London P13.
- Security/compliance/geo GTM: Customer Security Compliance Specialist E51; Software Engineer (Threat Intelligence), Cloudforce One E39; CEE customer-engineering leadership (London/Amsterdam) P8; Russian- and Hebrew-speaking customer engineers P12E22; Senior Named Account Executive, Saudi Arabia (Dubai) P9; Senior Sales Manager, CEE Digital Natives E25; Program Lead Strategic CDN Accounts GTM E21.
- People/ops: Workday Reporting Specialist P3 / Workday Analyst, People Operations E11; two Senior Recruiter roles E49E50; Logistics Operations Specialist E28.
Forks
- No cited evidence in this pack.
Releases
- cloudflare/computer @cloudflare/computer@0.7.0 — experimental
onInterrupt: "restart" | "replay"andctx.interrupt()/runtime.pending()/runtime.respond()for replayable runs awaiting approvals P17. - cloudflare/agents agents@0.30.1 — fixes
PiHarnessonWakenot resuming a sleeping task (compaction retry) P20. - cloudflare/cf cf@1.0.0-beta.21 — moves prompts/progress to stderr; records global CLI version in telemetry P18.
- cloudflare/workerd v1.20261024.1 (plus v1.20261023.1) — runtime fixes spanning SQLite stubs, JSG prototypes, streams, and Web Platform Tests P27E40.
- cloudflare/workers-sdk @cloudflare/vite-plugin@1.65.0 — drops
experimental.newConfig, Wrangler-config-only, Bun binding errors, local Workers runtime update P16. - cloudflare/workers-py workers-py-v1.17.9 E38.
- cloudflare/lol-html v3.0.2 — HTML parsing performance for tags split across chunks P19.
- cloudflare/networkquality-rs v0.4.0 — per-interval responsiveness progress P21.
- cloudflare/speedtest v1.15.0 — opt-in packet loss P15.
- Infra-as-code/schema: cloudflare/forge openapi auto-generated repeatedly from api-schemas P22P23P24P25P26; terraform-provider-cloudflare v3.35.1 E47; cf-terraforming v0.30.0 E52; tf-migrate v1.1.1 E54; cloudflare/nimbus docs packages E45E46.
Talking
- “Introducing Clef: Cloudflare's first open-source decision models, now on Workers AI” — product changelog framing decision models versus text generation, plus RL fine-tuning W1.
- Third-party Clef coverage: Apache 2.0 weights, Qwen3.8-27B / Qwen3.5-9B backbones, probability-over-options design W2W4.
- “Content Independence Day, one year on” — agentic-Internet economics; claims 2.5B genAI users in 3.5 years and collapsing open-web engagement P7.
- Post-quantum GeoKey Manager closed beta — harvest-now-decrypt-later defense, GA planned Q1 2027 P28; minimal outside traction (3 points / 0 comments) E1.
- Gartner SASE/SSE Visionary positioning — “first SASE platform to rein in MCP server sprawl, natively governing AI agents” P5.
- DDoS Threat Report H1 2026 — 1 Tbps attack surge, DNS floods, geopolitical vectors P4.
- Compliance/trust narrative — South Korea ISMS certification E60; civil-society/Project Galileo report ($100M donations milestone; 81.7% of malicious traffic was DDoS) P1.
- Engineering posts — Origin Range Requests rebuilt in Pingora, cutting R2 p99 latency in half E14.
Shipping
The most consequential shipped artifact is the Clef model family: @cf/cloudflare/clef and @cf/cloudflare/clef-flash are live on Workers AI, with Apache 2.0 weights on Hugging Face and a new RL fine-tuning service W1. Specs: Clef is fine-tuned on Qwen3.8-27B (~27.4B params, ~54.7 GB BF16) and Clef-flash on Qwen3.5-9B (~9.4B params, ~18.8 GB), both with 65,536-token context and vision input, priced $0.24 and $0.09 per million input tokens respectively W4. The models output per-option probabilities over typed questions rather than free text W2. Platform pricing sits at $0.011 per 1,000 “Neurons” with a 10,000-Neuron/day free tier W3.
Beneath the models, Cloudflare shipped agent/runtime plumbing: computer 0.7.0's interrupt/replay primitives P17, agents 0.30.1 recovery fixes P20, cf CLI beta ergonomics P18, and near-daily workerd runtime releases P27. Infra-as-code and schema tooling also churned (forge openapi generations P22P23P24P25P26, Terraform provider E47, cf-terraforming E52, tf-migrate E54).
Research themes
- Decision models as an alternative to autoregressive text for agent actions: Clef emits probabilities across a typed question schema, explicitly positioned “in the same family as Typesafe's Jev” W1W2.
- Reinforcement-learning fine-tuning as a product surface: a service to “tune Clef for your own workloads” W1.
- Machine-learning infrastructure and efficiency: signaled by Ensemble AI team additions W5.
- Agent runtime resilience: replayable/interruptible runs and workspace file access in @cloudflare/computer P17; agent harness recovery semantics P20.
- Post-quantum cryptography for geo-fenced key management (DLS closed beta, GA Q1 2027) P28.
The pack is thin on inspectable training detail: beyond the Qwen fine-tune lineage and weight metadata W2W4, there are no cited paper links, evaluation results, or safety/alignment research artifacts for Clef. Data-quality note: two page items are mislabeled in this pack — P2 is OpenAI's Hugging Face incident report and P6 is the Stanford AI Index 2026; neither bears on Cloudflare Workers AI, so both are excluded from the substantive analysis P2P6.
Hiring & scaling
Hiring concentrates on (a) the AI/agent layer and (b) global infrastructure, with commercialization roles layered on top. Austin is a named AI hub (Workers AI Engineer W6; Events Manager P11); Amsterdam/London anchor European customer engineering P8P12; Lisbon is a newer datacenter-ops location P13; Dubai serves Saudi/MEA named accounts P9; US datacenter ops span six metros P10. Infrastructure scale-out is visible in storage, network ops, load balancing, and deploy-at-scale roles E9E57E55E56, while agentic-commerce roles E42E43 and GTM automation P14 suggest early monetization of the agent platform. Recruiting capacity itself is scaling (two Senior Recruiters E49E50) alongside HR analytics (Workday Reporting Specialist P3).
Category implications
- Strategy: Cloudflare is bundling a first-party model layer (Clef + RL fine-tuning) with its edge network, framing Workers AI as the runtime for agent decisions rather than a generic inference endpoint W1W2. The “agentic Internet” narrative positions Cloudflare as the commercial control point for AI crawlers and content value P7.
- Infrastructure: The release cadence on workerd P27, computer P17, and agents P20, plus storage/network/deploy-at-scale hiring E9E55E56, implies sustained platform spend on edge runtime, storage, and networking — consistent with a neocloud deepening its own compute substrate.
- Product: Clef, Clef-flash, and RL fine-tuning convert Workers AI from inference resale into differentiated decision tooling for agents W1W4; post-quantum GeoKey Manager extends the same platform into regulated/sovereign accounts P28.
- Research: Evidence points to applied model work (fine-tuned open backbones, RL) rather than frontier pretraining W2W4, plus ML-infra efficiency hires W5. No cited evidence of pretraining or safety-alignment programs.
- Hiring: The mix of ML engineers E10W6, agentic-commerce engineers E42E43, and GTM automation builders P14 shows Cloudflare staffing the technical platform and its commercialization simultaneously.
- GTM: Security and compliance are the wedge — Gartner SASE/SSE Visionary positioning P5, ISMS certification E60, and post-quantum/data-sovereignty messaging P28 target enterprise buyers; demand marketing E59, events P11, and geo-account executives P9E25 extend the funnel.
Traction highlights
- Clef Hugging Face repo likes at time of writing: 457 (Clef) and 161 (Clef-flash) W4.
- Workers AI usage pricing: $0.011/1k Neurons, 10k Neurons/day free W3.
- DDoS mitigation scale: 23.2M network-layer and 29.64T HTTP DDoS requests in H1 2026; 935 attacks exceeding 1 Tbps P4.
- Cloudflare named the sole vendor “Visionary” in both 2026 Gartner SASE and SSE Magic Quadrants P5.
- Self-reported adoption claim supporting the agentic-Internet thesis: 2.5B active genAI users in 3.5 years P7.
- Project Galileo / Impact: $100M donation milestone; 81.7% of civil-society malicious traffic was DDoS P1.
- External attention to the post-quantum GeoKey Manager launch was minimal (3 points / 0 comments on HN) E1.