NewEnvironment snapshots now in public beta

Reproducible sandboxes
for every agent, every run.

Gradient gives your agents identical, byte-for-byte sandboxes. Spun up in under 100ms, destroyed in silence. One devfile, infinite warm clones, forkable in a single call.

Fleet · us-west · live
VC
verilog-copilot · synthesis agent
gradient.yamlverified · sha256:4f1a…
image: eda-base:2025.04
cpu: 4 vCPU
memory: 8 GiB
preload: [yosys, verilator, gtkwave]
Boot trace
00.000snapshot fetch18 MB
00.021vm bootfirecracker
00.038toolchain mountoverlayfs
00.051pdk cache warmsky130
00.072agent readyssh open
Fleet · 64 clones
48/64 ready
ReadyWarm poolBooting
Median boot
72ms
p99 ready
96ms
Forks today
38
Runs today
12.4K
What we're building
Opinion 01
Environments are the unit of reliability.
An agent is only as reproducible as the box it runs in. We ship the box, not a wrapper around someone else's.
Opinion 02
Cold starts are a bug, not a tradeoff.
Sub-100ms boot changes how you architect agent systems. One-shot runs become cheaper than long-lived daemons.
Opinion 03
A fork is faster than a Dockerfile.
Start from any image, layer new libraries, binaries, or tools on top, snapshot it. Every fork is a first-class environment — versioned, reproducible, cloneable.
The platform

Three primitives. One agent-native runtime.

Everything an agent needs to run a task — reliably, repeatedly, without a platform team.

01 / SNAPSHOTS
DEVFILE

One devfile, infinite identical clones.

Declare the environment once. Every agent run gets a byte-for-byte copy — same deps, same caches, same kernel. No more “works on my VM.”

  • Content-addressed VM snapshots
  • Hydrated package caches (pnpm, pip, cargo)
  • sha256-verified at boot
02 / RUNTIME
Docker3.2sCold Start750msGradient62ms

Sandboxes that boot in under 100ms.

Firecracker microVMs with pre-warmed page caches. Your agent spawns a fresh, isolated environment faster than most CI systems pull an image.

  • p99 boot under 140ms
  • SSH-only, no shared kernel
  • Regional placement for low latency
03 / FORKS
BASEubuntu+ yosys+ pandoc+ cuda

Fork any image. Add what you need.

Start from a base — ubuntu, cuda, node, or your own. Layer on libraries, CLI tools, dependencies. Snapshot the result. Every fork is a first-class environment, ready in milliseconds.

  • Fork-and-extend any image
  • Install binaries, toolchains, shared libs
  • Versioned — roll forward or back in one call
Live on the platform

Every fleet, right-sized automatically.

Because every environment is reproducible, we know exactly what it needs. Watch real fleets being tuned in real time.

fleet-telemetry.live
1 / 3
Currently tuning
chip-design copilot · synthesis fleet
127 end-users · 1,240 agents live
Observed
4.2 GB used
Right-sized to
6 GB allocated
Was
16 GB
■ Actual usage■ Right-sized allocation■ Previous allocationmax 16 GB
Median boot
72ms
Cost / agent-hour
$0.036
was $0.098 · −63%
Active forks
38
versioned snapshots
Get started

One gradient.toml.
Run your agents in identical machines.

Install the CLI, declare your environment, point your agent at a sandbox.