Divergent Labs

The compute platform for agents

Give your agent superpowers, with sandboxes to work in, models built for the job and supercomputers to train on.

Get started

All in an interface your agents understand

One CLI for sandboxes, models and GPUs. Great documentation, plain commands, machine readable output agent can act on.

~/very-cool-classifier
# Run a training run in a sandboxed environment with a $5 budget
$ div run --gpu h100 train.py::run --max-spend 5.00

# Pull and push anything to super fast block storage
$ div storage get training-volume-1 log/training.jsonl

# Stream logs your agent can monitor and act upon in realtime
$ div logs training-run-4 -f

# Give every agent its own machine, with a browser built in
$ div sandbox create --size standard

Divergent Labs gives the most ambitious teams the tools to build their own agentic software factory. We give agents everything they need to research, build and test.

Sandboxes for Agents

Every agent gets a machine to build, test and browse on, for as long as the job takes.

Agenttrip
Cheapest nonstop from Amsterdam to SF, Oct 20 to 27? Mornings if possible.
div sandbox inputtrip type 'San Francisco'✓ typed
div sandbox a11ytrip do n84 press✓ Cheapest
div sandbox a11ytrip do n112 press✓ KLM 10:25
KLM at 10:25, €512 round trip. Prices are low for these dates, so now is a good time to book.
Book it. Window seat.

Supercomputers for Agents

Your agent trains, evaluates and ships its own models, on GPUs you only pay for while they run.

Agentbudget $2,000
Fine-tune a 70B open model on our 38k annotated contracts. It has to beat our frontier-model baseline on clause review. Budget is $2,000.
LoRA on 4 nodes of 8× H200, evaluating against your held-out set every 600 steps.
div run--gpu h200 --gpus 32 finetune.py::lora --max-spend 1500✓ $1,128
div eval runclause-review --model contracts-70b:run-7 --gpu h200 eval.py✓ 87.2% vs 81.5%
Run 7 beats the baseline by 5.7 points. Eval loss bottomed out at step 5,400, so I kept that checkpoint.
div model publishcontracts-70b:run-7 --volume checkpoints --path contracts-70b/run-7✓ private
Deploy it on 2× H200 behind an endpoint
contracts-70b · run-7done
traineval
$ div logs run-7 -f
step 4500 loss 0.522
step 4800 loss 0.559 eval 0.667
step 5100 loss 0.518
step 5400 loss 0.513 eval 0.657
step 5700 loss 0.512
step 6000 loss 0.545 eval 0.740
step6,000
best eval0.657
spend$1,128

Pricing

Pay for what your agents use, billed by the minute.

Get started today

Nothing to install. Your SSH key is your account.

~
# Sign up with just your SSH key
$ ssh div.io

# Create a sandbox
$ ssh div.io new --name agent-box

# Open a shell on it
$ ssh agent-box@div.io

Give your agents a lab of their own.

Get started