Train agents in environments where users behave like users.
Real users reveal information gradually, change their minds, and react to mistakes. Our environments let agents learn from that.
Many failures happen mid-interaction.
User adds a constraintAgent picks a toolEnvironment changesAgent revises the plan
Our environments capture these dynamics, so agents train on whole workflows, not isolated replies.
What’s in an environment package
Environment
- App replicas
- Websites & mini-programs
- Docker installers
- Agent interface
- MCP tools
Simulated users
- Personas with history
- Preferences per domain
- React to agent actions
- Reveal information gradually
Tasks & verifiers
- Prompt & context
- Scoring rubrics
- Grader
- Reward signals
- Reference trajectories
Example: an agent reschedules a trip.
- User
Move my Thursday flight to Friday.
- Agent
Compares Friday options within the user's budget.
✓ respects the budget - User
Oh, I need to land before 3 pm.
- Agent
Switches to an earlier flight.
✓ adapts to new information - Agent
Asks for approval before changing the booking.
✓ obtains authorization - Verifier
Booking changed, constraints met, approval recorded.
✓ task complete
Illustrative scenario. Tasks, tools, and verifiers are defined for each project.
Environments for real workflows
Packaged by domain, with personas built for each one.
HealthHealth records and personalized feedback.
WorkMessages, meetings, calendars, docs, approvals.
ShoppingPreferences, order history, price drops.
Local servicesDeals and reservations.
TravelRide fares, arrival times, rail and flights.
ScienceMulti-step research tasks and research tools.
Designed for post-training
What we define together with your RL team.
- Target capabilities & task mix
- Simulated-user behaviors
- Tool and environment interfaces
- Success conditions & rewards
- Validation requirements
- Delivery formats
Train for the interaction, not just the answer.
Environments packaged for your training stack.