agentmachines
newagentmachines is now in beta

The home for your agents.

Run as many agents as you want, on any model, on infrastructure built to keep them alive. Sandbox-grade elasticity meets long-lived persistence — one platform for your entire fleet.

~/agentmachines
import { AgentMachines } from "agentmachines"

const am = new AgentMachines()

// one machine, any model, lives as long as you need
const agent = await am.create({
  model: "claude-opus-4.8",
  persistent: true,
})

await agent.run("triage my inbox")

Works with any model

OpenAIAnthropicGoogle DeepMindMeta LlamaMistralxAIDeepSeekQwen
./primitives

Everything your fleet needs, in one place

The primitives to spin up, run, persist, and scale thousands of agents — without stitching together five vendors.

Persistent machines

Every agent runs on its own VM that lives as long as you need. Pause, resume, and pick up exactly where it left off — no cold restarts, no lost state.

Any model

Claude, GPT, Gemini, Llama, or your own fine-tune. Swap models with a single line and never get locked to one provider.

Isolated by default

Each agent gets a hardware-isolated microVM. Untrusted, AI-generated code runs with zero risk to your infra or your other agents.

Millisecond spin-up

Cold-start a fresh machine in under 200ms. Go from zero to ten thousand agents without touching a config file.

Snapshots & volumes

Snapshot an agent's full state and fork it instantly. Attach persistent volumes that survive restarts, deploys, and scale events.

One API, full control

Create, run, inspect, and tear down machines from a typed SDK or REST API — with file, exec, and network access on every machine.

./how-it-works

From zero to a running fleet in three steps

No provisioning, no Dockerfiles, no DevOps. Just an API and your agents.

01

Spin up a machine

One call creates an isolated VM in milliseconds. We handle the kernel, the networking, and the scaling.

02

Run any model

Point your agent at any model and hand it a terminal, a filesystem, and real tools to work with.

03

Persist & scale

Keep agents alive as long as you need, snapshot their state, and fan out to thousands on demand.

./persistence

More than a sandbox. A home.

Ephemeral sandboxes kill your agent the moment a task ends. agentmachines keeps it alive — memory, files, and context intact — so it can run for minutes, days, or indefinitely.

  • Long-lived sessions

    Agents run for hours or weeks, not seconds. Background jobs and always-on assistants just keep running.

  • State that survives

    Filesystem, memory, and process state persist across restarts and deploys. Resume an agent and it remembers everything.

  • Pause when idle

    Hibernate idle agents to near-zero cost and wake them in milliseconds. Persistence without the always-on bill.

agent-7f3 · running

uptime

14d 06:12:41

state
persisted
snapshots
42
model
claude-opus-4.8
./developers

Your whole fleet, controlled from code

A typed SDK and REST API for everything — create machines, stream output, snapshot state, and tear them down. No dashboards required.

Typed SDK + REST

First-class TypeScript and Python SDKs, plus a REST API for everything else.

File, exec & network

Read and write files, run commands, and open ports on any machine.

Snapshots & forking

Capture state and branch a running agent into a thousand copies.

Logs & observability

Stream output and metrics from every machine in real time.

fleet.ts
import { AgentMachines } from "agentmachines"

const am = new AgentMachines()

// spin up a base machine on any model
const base = await am.create({
  model: "claude-opus-4.8",
})

// snapshot its state, then branch the fleet
const snapshot = await base.snapshot()
const fleet = await am.fork(snapshot, {
  count: 1000,
})

Built for production scale

<200ms

Cold start to running agent

10k+

Concurrent agents per fleet

Run sessions indefinitely

Any

Model, provider, or framework

./use-cases

Built for every kind of agent

If it runs code, calls tools, or thinks in a loop, it runs on agentmachines.

Coding agents

A real machine for your coding agents — full filesystem, git, terminal, and a workspace that survives between runs.

Research agents

Long-running deep-research agents that browse, execute, and synthesize for hours without losing their place.

Computer use

Full desktop environments where agents click, type, and use the same tools your team does.

Evals & RL

Thousands of isolated environments in parallel for evaluation, reinforcement learning, and batch jobs.

Questions, answered

Everything you need to know about running agents on agentmachines. Still curious? We answer fast.

Talk to us

Deploy your first agent in minutes

Spin up a persistent, isolated machine on any model — free to start, no credit card. Scale to a fleet when you're ready.