Better agents, on data that stays yours.

Tell us what the agent should do. Our agents set up the workflow, connect your tools and systems, and test it on real tasks before it goes live.

• Templates for common agent workflows
• Connects to your tools, APIs and data sources
• Tested on your own tasks before launch
• Runs on the models you choose, frontier or open

Point your agent at one endpoint, without a rewrite. It sits in front of the frontier models you use today, open models, and later your own.

• Works with the API format your agents already use
• Runs in our cloud or inside your VPC
• Shift traffic between models gradually, and roll back in one step
• A full copy of every run, kept in storage you control

Our agents read your agent’s runs, find where it fails, and propose fixes to prompts, tools and workflows. Every change is checked on your own tasks before it ships.

• Failures found and sorted by cause
• Fixes tested against the version you run today
• You approve every change
• When the runs add up, a model of your own

Every model call, tool call and result in one place. Find out why a run failed, compare models on your own tasks, and turn the runs that worked into training data.

• Full traces of agent runs, searchable and replayable
• Evals on your own tasks, for any model or agent version
• From a set of good runs to a training dataset in one step

Traces stay in the environment you choose: our cloud, your VPC or your own data center. Only the people you choose can see them, and they train only your model.

• Stored in your environment, never pooled with other customers
• Access limited to the people you choose
• Never resold, never used to train anyone else’s model

Why keep your data →

We help you start from an open model and train it on what your agent actually does. First on examples of good runs, then by letting the model try tasks in sandboxes and rewarding the ones it gets right.

• Supervised fine-tuning on your traces, then reinforcement learning with verifiable rewards
• Evals against the model you use today, on your own tasks
• Trained on your data and open models, never on outputs of closed frontier models
• You own the weights and can run them anywhere

Agent traffic is bursty, long-context and full of tool calls. We serve open models, and the models we help you train, with steady latency under that load.

• Leading open models, ready to serve
• Caching and routing tuned for multi-turn, tool-heavy sessions
• Committed token capacity or dedicated inference capacity

Agents write code, run commands and open browsers. Each sandbox is an isolated environment that starts fast and disappears when the task is done. During training, thousands run at once, next to the GPUs.

• Isolated environments for code, shells and browsers
• Snapshot and fork a running environment
• Control what each sandbox can reach on the network
• Thousands in parallel for training, in the same data center as the GPUs