RL environments

RL environments.Sandboxes where agents learn by doing.

Our RL platform gives agents and LLMs safe sandbox environments to act in, with tasks and reward signals written by the same vetted experts behind our data.

RL environments · finance-ops-v2SL
Episode 0142
SandboxedCompleted
Task · written by a finance expertReconcile the March ledger against the bank statement and flag mismatches.
01open_fileledger_march.csv
02querybank_statement WHERE month = 3
03compare41 rows matched · 3 mismatched
04writeflags.json
05submit3 flagged items with reasons
Reward
Correct
Follows policy
Efficient
Reward over training
Illustrative interface

What the platform does

Agents and LLMs take actions in sandboxed environments, get scored on the outcome, and improve from that signal. The sandbox keeps every action contained, so a model can try, fail and recover without touching real systems.

  • Sandboxed environments for agents and LLMs to act in
  • Tasks written by experts in the field the agent works in
  • Reward signals from automatic checks, expert graders, or both
  • Episode logs you can replay, audit and use for evaluation

Why experts write the tasks

An environment is only as good as its tasks and its rewards. A reward that a model can game teaches the wrong thing. Our experts write tasks that look like real work in their field, and design rubrics that reward the right outcome, not a shortcut to it.

Training and evaluation

The same environments serve both purposes. Use them to train with reinforcement learning, or hold tasks out to evaluate an agent before release.

Questions, answered.

Can we bring our own agent or model?

Yes. Your agent or model runs against our environments, and the scope, tasks and rewards are agreed in the written spec for the engagement.

Can you build an environment for our domain?

Yes. Tell us the work the agent should do. We source experts in that field to write tasks and rewards, then run a pilot before scaling.

Every safe model
has an expert behind it.