RL ENVIRONMENTS / VERIFIERS / EVALUATION
Gantovar builds RL environments for teams training and testing AI agents: sandboxed copies of real software, tasks with a written definition of done, and verifiers the agent can't reach.
Start with a callHOW WE BUILD
If a verifier can be satisfied without doing the work, your reward signal teaches the wrong thing. We write the check first.
State is seeded and resets are tested, so a change in score means a change in the agent and not in the environment.
Before handoff we attempt the shortcuts an agent would find: reading the answer, editing the check, ending early.
WHAT WE BUILD
Real software, in a box. Sandboxed versions of the tools your agents work in, packaged as tasks with seeded state and tested resets.
A score you can believe. Automated checks kept outside the agent's reach, reviewed against the shortcuts agents actually take.
Know if it got better. Held-out tasks, repeated runs and reports that make one model comparable with the next.
METHOD
A call, then a written list of tasks, each with a definition of done and what counts as cheating.
Sandbox, seeded state, tasks and verifiers, delivered in small reviewable batches.
We attempt the shortcuts, fix what breaks and record what remains.
Versioned containers, documentation and a run report, set up to run on your side.
QUESTIONS
Gantovar is a new company and has no published client list or case studies yet. Every engagement starts with a written scope and acceptance criteria, so you can judge the work directly.
It depends on the number of tasks and how hard the software is to reproduce. We quote after the scoping call.
Some work may be delivered with specialist partners. If so, we say who and for what during scoping.
That's the default. Environments ship as containers so that your data and agents can stay on your side.
Tell us what the agent should do and how you'd know it worked. We'll reply with whether we can help.
hello@bestrlenvironments.com