Cloudflare Agent Cloud
by Cloudflare
Cloud infrastructure rebuilt for agents as the primary workload, not humans
Cloudflare Agent Cloud is the company's umbrella platform for building, hosting and observing AI agents on its global network. It spans the Agents SDK and Project Think harness, persistent Sandboxes and Artifacts for compute, Agent Memory and AI Search, a voice pipeline, and agent-level tracing — all built on Workers and Durable Objects.
Cloudflare Agent Cloud is the company's positioning of its developer platform as what it calls Cloud 2.0 — infrastructure designed for a world where agents, not humans, are the primary workload. The argument is architectural: traditional cloud services assume one app serving many intermittent human users, whereas a single worker may run several agents in parallel around the clock, creating different requirements for speed, structured data access and system integration. The foundation is the Agents SDK, which runs stateful agents on Workers and Durable Objects with durable identity, local SQL storage, WebSocket real-time connections, scheduled work and recoverable execution through Fibers. Project Think is the opinionated harness layered above it for teams that do not want to assemble their own. Agents Week 2026 filled out the surrounding platform: Sandboxes reached general availability as persistent isolated environments with shell, filesystem and background processes; Artifacts added Git-compatible versioned storage; Durable Object Facets let workers instantiate isolated SQLite databases; Workflows was rearchitected for 50,000 concurrency; and Cloudflare Computer, an open-source runtime in early preview, routes execution across isolates, containers and browsers over a shared SQLite filesystem with the goal of needing a container for under 10% of an agent's work. The toolbox adds Agent Memory, AI Search, Browser Run, a real-time speech pipeline over WebSockets, an Email Service in public beta, and an AI Platform inference layer spanning 14+ providers. Security includes Cloudflare Mesh, Managed OAuth for Access via RFC 9728 and scoped API tokens. Cloudflare Agents, announced 4 August 2026, adds agent-level tracing, session replay and trace waterfalls.
The platform engineering team building customer-facing or internal AI agents that need persistent state, global low latency and real observability, without standing up and operating their own agent infrastructure.
Stateful agents with durable identity, memory, sandboxed compute and per-agent tracing deployed globally on one platform, with no servers to manage.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Freemium, Usage-based, Subscription, Contact for pricing
- Target Market
- CTOs, Enterprise Developers, Platform Engineers, AI Engineers
- Deployment
- Cloud-only, Edge-first, API-based
- Founded
- 2009
- Headquarters
- San Francisco, United States
- Team Size
- 500+
- Customers
- Not disclosed for the agent platform specifically
Key Features
- ✓Agents SDK
Runs stateful agents on Workers and Durable Objects with durable identity, local SQL storage, sessions, routing and real-time WebSockets.
- ✓Durable execution with Fibers
Recoverable execution means a long-running agent survives failure and resumes rather than losing its work entirely.
- ✓Sandboxes (GA)
Persistent isolated environments giving agents a real shell, filesystem and background processes, with Outbound Workers providing zero-trust egress control.
- ✓Agent tracing and session replay
Captures every model call, tool execution, approval event and subagent call, and replays whole conversations without re-running the agent.
- ✓Agent Memory and AI Search
A persistent memory service plus a hybrid-retrieval search primitive, so agents keep context and query knowledge without bespoke infrastructure.
- ✓Cloudflare Computer
An open-source runtime routing work across isolates, containers and browsers over a shared SQLite filesystem to cut container dependence below 10%.
- ✓AI Platform inference layer
A unified inference abstraction across 14+ model providers, so agents are not hard-wired to one vendor's API.
- ✓Voice pipeline and Email Service
Real-time speech-to-text and text-to-speech over WebSockets plus an email channel in public beta for multi-channel agents.
Capabilities
Use Cases
- •Customer-facing conversational agents
Deploy stateful chat agents globally with persistent memory and real-time WebSocket connections close to every user.
- •Debugging agent failures in production
Agent tracing surfaces why an agent returned HTTP 200 while failing internally, showing tool choices and token consumption.
- •Agents that execute untrusted code
Sandboxes provide isolated shell and filesystem access, with Outbound Workers proxying egress so agents cannot reach arbitrary destinations.
- •Voice agents over telephony or web
The real-time speech pipeline handles streaming transcription and synthesis for agents that must hold a spoken conversation.
- •Long-running agentic workflows
Rearchitected Workflows support 50,000 concurrency with durable execution for orchestration that spans hours or days.
Ideal For
Best For
- ✓Building stateful AI agents that need durable identity, persistent memory and recoverable execution across long-running sessions
- ✓Teams needing agent-level observability — which tools were chosen, tokens consumed, and how subagents performed — rather than HTTP status codes
- ✓Giving agents sandboxed compute with a real shell, filesystem and background processes for code execution and tool use
- ✓Multi-channel agents reachable over chat, voice, email, Slack and webhooks from a single runtime
- ✓Globally distributed agent workloads where per-region latency matters and infrastructure management is unwanted
Not Ideal For
- ✗Teams committed to a single hyperscaler's managed AI stack, since this means adopting Cloudflare's runtime model of Workers and Durable Objects rather than standard containers or VMs
- ✗Workloads needing long-running heavyweight compute, given Workers cap invocation duration at five minutes and the model favours isolates over containers
- ✗Anyone wanting to build on Cloudflare Computer today — it is in early preview and Cloudflare states it is suitable only for experiments, exploration and prototypes
- ✗Buyers who need every component production-hardened now, as much of the Agents Week surface shipped as beta or preview rather than GA
Integrations
Deployment
Market & Ratings
Not disclosed for the agent platform specifically
Market Analysis
Pros
- ✓Genuinely broad primitive coverage — compute, storage, memory, search, voice, email, browser and multi-provider inference on one platform
- ✓Durable Objects provide per-agent state with co-located SQL storage, a structural advantage for stateful agents over stateless container platforms
- ✓Agent-level observability addresses the real failure mode, where an agent returns HTTP 200 while failing internally
- ✓No data transfer or egress charges, which matters for agents making heavy outbound tool and API calls
- ✓Global deployment with automatic scaling and no infrastructure to manage, plus a generous free tier for evaluation
Cons
- ✗Much of the announced surface is beta or preview rather than GA — Cloudflare Computer is explicitly early preview and stated to suit only experiments and prototypes, while independent coverage found Agent Memory and Artifacts in private beta, AI Search in open beta and Flagship in closed beta
- ✗Adopting it means adopting the Workers and Durable Objects programming model, which is materially more opinionated than containers and hard to port elsewhere
- ✗A five-minute maximum invocation duration constrains heavyweight compute, pushing long tasks into Workflows or containers
- ✗Observability moves from free to metered on 1 October 2026, so beta-era cost assumptions will change for anyone relying on tracing at volume
- ✗Framework support for tracing is currently limited to Think, Flue and the AI SDK, with broader OpenTelemetry support still in development
- ✗The sheer breadth announced in one week is hard to evaluate — press coverage was largely descriptive, with no independent analyst benchmarking against AWS, Azure or Google agent platforms
Pricing
Workers Free
$0
- ✓100,000 requests/day
- ✓10ms CPU time per invocation
- ✓Durable Objects 100,000 requests/day on SQLite backend
- ✓D1 5GB storage
- ✓Workers KV 100,000 reads/day
Workers Paid
From $5/mo
- ✓10M requests/month included, then $0.30 per additional million
- ✓30M CPU-milliseconds/month included
- ✓5-minute max invocation duration
- ✓Durable Objects 1M requests/month
- ✓Containers billed per GiB-second and vCPU-second
- ✓No egress or bandwidth charges
Enterprise
Contact for pricing
- ✓Custom contracts and commitments
- ✓Advanced security controls
- ✓Enterprise support
There is no separate Agent Cloud SKU — you pay ordinary Cloudflare developer platform rates. Workers Free covers 100,000 requests a day; Workers Paid starts at a $5/month minimum and includes 10M requests plus 30M CPU-milliseconds, then bills $0.30 per additional million requests and $0.02 per additional million CPU-milliseconds. Durable Objects, which back agent state, include 1M requests and 400K GB-seconds monthly. Containers powering Sandboxes bill at $0.0000025 per GiB-second of memory and $0.000020 per vCPU-second. Cloudflare Agents observability is free during beta through 30 September 2026; from 1 October it folds into Workers Observability pricing at 200,000 events daily free with 3-day retention, or 20M events monthly on paid with $0.60 per additional million and 7-day retention. Notably, there are no data transfer or bandwidth egress charges at all.
Security & Compliance
Sources
This page was written from 7 sources, 5 on domains other than developers.cloudflare.com.
- 1.blog.cloudflare.com — agents week in review
- 2.blog.cloudflare.com — agents on cloudflare
- 3.developers.cloudflare.com — agentsvendor
- 4.developers.cloudflare.com — pricingvendor
- 5.infoq.com — cloudflare computer agents
- 6.startuphub.ai — cloudflare plans for the agent cloud
- 7.vktr.com — cloudflare agents week
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
WEKA NeuralMesh
Microservices storage and memory fabric built for AI training and inference at exabyte scale
Tsuga
Bring-your-own-cloud observability that keeps telemetry, and its cost, inside your own AWS account
OpenObserve
Open-source, Rust-based observability on object storage — logs, metrics, traces and LLM telemetry in one binary
Coralogix
AI-native observability that queries logs, metrics and traces in place — index-free, in your own S3 bucket