What is Promigence? Sandboxes for AI agents, explained

The name#

It is spelled P-R-O-M-I-G-E-N-C-E. Search engines that have not met the word yet like to correct it to "prominence", so if you searched for Promigence and got pages about prominence, you typed it right.

The product is Promigence. The company behind it is Promigence AI, Inc., a Delaware corporation based in San Francisco, and its home is www.promigence.ai.

Why an agent needs a sandbox#

An agent runs code nobody reviewed. A model wrote it, and anyone who can type into your product can steer that model. Run the code on a laptop or on your own servers and a mistake, or a prompt someone planted, has the same reach as the account it runs under. A sandbox gives each task a computer of its own, with its own files, network rules and spend cap, so whatever goes wrong stays inside it and ends when the task ends.

What teams run on it#

Take coding agents. A coding agent gets a copy of your repository that is already set up, works on a task in it, and hands back the change. Running a coding agent in a sandbox takes one command.

CI jobs fit too. A workflow on a supported hosted CI service moves onto Promigence sandboxes with a one-line change and is billed per second; the guide is a CI runner in two lines.

Then there is code your users' AI writes. If your product lets a model write and run code for the people using it, give each user or session a sandbox of its own, with the network off or limited to an allow-list, plus a spend cap and a timeout. Three guides cover that shape: running AI-written code safely, a sandbox for every user of your product and a code-execution tool your agent can call.

Evals and reinforcement learning need many identical environments at once. Build the environment once, then start a fresh copy for every eval task or RL episode, a thousand at once if the work calls for it. The validity report that comes back separates environment failures from task failures, so a broken environment never counts against your agent, and a snapshot named by a hash of its contents lets a result cite the exact environment it ran in. Snapshot after an install, then start ten copies shows the pattern, and the guides page lists the eval harnesses and agent frameworks that have a guide so far.

Agents inside apps get the same treatment: a sandbox per user that can sleep when the user stops and wake on their next call. A paused sandbox is not billed by the second and is kept for 30 days. The same pattern serves security testing: give a testing agent a fresh copy of your own application for every attempt with outbound access limited to what the test allows.

How using it looks#

There are three steps, from the command line or the SDKs.

First, build a snapshot of your project. You name a base image with your toolchain, the setup steps (a clone of your repository at a pinned commit, then the install) and a verify command, which is your own test command. If the verify command does not exit zero, nothing is stored.

Second, price the run before it starts. promigence run quote gives a ceiling, as if every sandbox ran for its whole timeout, and --cap turns that into a hard limit: a run quoted over its cap is refused before anything starts, and a sandbox that reaches its cap is paused with its files kept. The quickstart's example prices 20 small sandboxes for up to 15 minutes each at $0.83 at most.

Third, start the copies. One command starts as many sandboxes as you ask for, runs your command in each and prints the validity report. A new account runs 20 at once.

promigence snapshot create \
  --image <your base image> \
  --setup "<clone your repository at a pinned commit>" \
  --setup "<install your dependencies>" \
  --verify "<your test command>" \
  --tier small --alias myapp --wait --wait-timeout 10m

promigence run quote --snapshot myapp --count 20 --timeout 15m
promigence run --snapshot myapp --count 20 --cap 1 -- <your command>

Billing is per second, at $0.0504 per vCPU-hour plus $0.0162 per GiB-hour. A small sandbox (2 vCPU, 4 GiB) comes to $0.17 an hour, and one that runs for ten minutes is billed for ten minutes. Paused time costs nothing.

Numbers you can check#

The home page publishes its measurements with the conditions attached. Two matter on a first look. A sandbox is ready to run its first command 8 ms after it is created, at the median (create to first command, 2 vCPU / 4 GiB, n = 100, unloaded). Start 100 at once and even the slowest 5% finish in under a second: 0.94 s at the 95th percentile, with every run completed (n = 400).

Security#

No sandbox can see, reach or address another, yours or anyone else's, and that boundary is tested by attacking it. A sandbox can have full outbound access, none, or an allow-list of the destinations it needs. Secrets are stored once, given only to the commands that name them, and never written into a snapshot you build. Your code is not used to train models.

How to get started#

Leave your work email under Get started on www.promigence.ai, or write to support@promigence.ai. New accounts start with $50 of free credit for 30 days: $10 claimed in the browser with no card, and $40 more once a card is added.

From there, the quickstart goes from nothing to 20 verified runs in about ten minutes, most of it the first build. To see your own workload first, send a repository and its command for the free validity report: 1,000 runs, the timings, the failures and an exact price.

Sources#

Every page below was read on 19 Oct 2026.

Updated 2026-10-19