Platform

Execution Primitive
Micro-VM Compute Built for AI Agents

Sandboxes runs your agents in real Linux devboxes on a custom bare-metal hypervisor. Boot in under two seconds from cached blueprints, snapshot and branch state, suspend devboxes between work to burn no compute, and mount code, agents, files, or S3 with the same primitive.

why runloop

Devboxes, Blueprints, Snapshots, Tunnels, and Mounts

Sandboxes is the Execution primitive, the foundation every other Runloop primitive runs on. A devbox is a hardware-isolated micro-VM with a full Linux environment: filesystem, shell, networking, and persistent state. Blueprints define and pre-build the environment as a reusable image. Snapshots checkpoint disk state for branching and restore. Tunnels expose ports publicly or behind a token. Mounts attach code, agents, files, S3 objects, or Axon brokers.

Text button?
SEACTION FEATURE

Feature Block Alternating

Run structured benchmarks and compare models before deploying to production.

Blueprints

Build the Environment Once, Boot Devboxes From It Forever

A blueprint is your devbox environment as code: a Dockerfile, system setup commands, code mounts, build args, secrets, named contexts, and network policies. The platform builds it (QUEUED, PROVISIONING, BUILDING, SUCCESSFUL) and stores the image, so every devbox launched from that blueprint boots from a warm cache.

Sub-second cold start — apt, pip, clone, build all happen once at blueprint build time, not on every devbox.
Versioned and inheritable — blueprints reference base blueprints; each update is a new immutable image.
Build from inspection — create blueprints from an existing devbox's installed state, not just from a Dockerfile.
Same blueprint, every primitive — agent workloads, benchmark scenarios, and production runs all launch from the same image.
Find Out More
Benefit image.
dashboard card UI mockup
Snapshots

Branch, Resume, and Fork Devbox State

A snapshot captures the full disk state of a devbox. You can launch a new devbox from any snapshot, fork the same checkpoint into parallel devboxes, or roll a devbox back. Snapshots flow through their own state machine (QUEUED, PROVISIONING, SNAPSHOTTING, SUCCESSFUL) and are first-class through both sync and async APIs.

Parallel exploration — snapshot a working agent, then fork the same point into many devboxes to try different approaches.
Training-loop efficiency — restore from snapshot at each fine-tuning step instead of rebuilding environments from scratch.
Rollback to any point — promote a working snapshot to production, roll back instantly if a regression appears.
Sync or async — small snapshots return synchronously; large ones run in the background and notify on completion.
Lifecycle

Suspend Between Work, Resume on Demand, Burn No Compute

A devbox is not just an ephemeral container. It moves through a state machine, provisioning, initializing, running, suspending, suspended, resuming, shutdown, that you control through the lifecycle API. A suspended devbox holds its full state and costs nothing in compute. Combine with wake_on_http and a tunnel call resumes the devbox automatically.

Zero compute while suspended — pay for running devboxes, not waiting ones.
State preserved — resume picks up disk and process state exactly where they left off.
wake_on_http — tunnels can resume a suspended devbox on first request; http_keep_alive holds it warm under load.
Multi-week pauses — long-running agents and HITL workflows can pause for as long as they need.
Find Out More
Dashboard showing last month spending
prduct map

Flexible infraestructure for any AI Workload

Runloop is the batteries included platform designed for building and optimizing AI-driven software engineering agents.

FAQ'S Only one section

Everything You Need to Know

We’re dedicated to solving the complex challenges of productionizing AI for software engineering at scale.

How easy is it to integrate Runloop with existing AI development pipelines?
What makes Runloop's AI code execution infrastructure enterprise-grade?
How does Runloop ensure safe and secure code execution for AI agents?