Comparison

A VM-first alternative when agent workloads outgrow edge runtimes.

Cloudflare Workers is excellent for request-driven code at the edge. But many agent workloads need something different: persistent state, full system access, longer execution times, and — critically — a signup flow that an AI agent can complete without human help.

Agent Cloud focuses on that gap. Instead of edge functions with millisecond cold starts, you get full Linux VMs that an agent can provision, configure, and manage through a simple REST API.

Side-by-side comparison

CapabilityCloudflare WorkersAgent Cloud
RuntimeV8 isolates (JS/Wasm)Full Linux VMs (any language, any tool)
SignupHuman account with email and dashboardSingle API call, no email required
Agent self-provisioningNot supportedCore design principle
Execution time10ms-30s (depending on plan)Unlimited (VM runs until stopped or expired)
System accessSandboxed V8 — no filesystem, no shellFull root access, SSH, install anything
Persistent stateKV, D1, R2 (separate services)Local disk on the VM
Free tier100K requests/day1 micro VM for 72 hours, no card required
Billing modelPer-request and duration pricingSimple plans, human-approved Stripe checkout
Best forLow-latency request routing, edge logicAgent tasks, sandboxes, longer-running jobs

When to use Cloudflare Workers

Workers is the better choice when you need:

  • Sub-millisecond cold starts for request-driven HTTP handlers
  • Global edge distribution close to end users
  • High-throughput, low-latency API routing
  • Tight integration with Cloudflare's CDN, WAF, and DNS

If your workload is request-driven, stateless, and needs to run at the edge, Workers is hard to beat.

When to use Agent Cloud

Agent Cloud is the better choice when:

  • An AI agent needs to provision its own compute. Workers requires a human to set up an account, configure wrangler, and deploy code. Agent Cloud lets the agent do all of this via API.
  • The workload needs full system access. Installing packages, running shell commands, compiling code, accessing the filesystem — these are natural on a VM and impossible in a V8 isolate.
  • Tasks run for minutes or hours. Agent research jobs, data processing, multi-step tool use — these exceed Workers' execution limits.
  • You need an isolated sandbox. Code execution environments, untrusted workloads, and agent orchestration prototypes benefit from full VM isolation.

They work well together

This isn't purely either/or. A common pattern is using Workers for lightweight API routing and Agent Cloud VMs for heavier agent tasks behind the API. The agent provisions a VM when it needs persistent compute and tears it down when the task is done.

Try the quickstart to see how an agent provisions a VM in four API calls.